Claude Sonnet 5

Claude Sonnet 5 is a large language model developed by Anthropic, released on June 30, 2026 as the mid-tier Sonnet model of the Claude lineup and, at launch, the default model for Free and Pro plans on claude.ai. ^[1][4] It is the successor to Claude Sonnet 4.6, sits above Claude Haiku 4.5 and below the Opus, Fable, and Mythos models, and at launch Anthropic described it as "the best combination of speed and intelligence." ^[2] Anthropic says its performance is "close to that of Opus 4.8, but at lower prices." ^[1] Sonnet 5 launched at $2 per million input tokens and $10 per million output tokens, originally announced as introductory pricing through August 31, 2026. On August 10, 2026 Anthropic made those rates permanent and cancelled the planned September 1 increase to $3 / $15. ^[1][35]

On September 28, 2026 Anthropic released Claude Sonnet 5.5 at the same list prices, and Sonnet 5 became a legacy model. As of September 29, 2026 it remains available on the API, with retirement scheduled for no sooner than June 30, 2027, and it has a new role as the fallback model that answers certain cybersecurity and AI-development requests that Sonnet 5.5's safeguards block. ^{[21][22][23][26]}

The model is more interesting to developers than a normal point release, because three API behaviors changed at once and a fourth change is invisible in the rate card: Sonnet 5 ships with a different tokenizer that turns the same text into approximately 30% more tokens than Claude Sonnet 4.6 did, and closer to 1.4x on English prose. ^[2][11] Sonnet 5's per-token prices are lower than Sonnet 4.6's, but the higher token count means per-request costs do not fall in direct proportion. ^[2][6]

What is Claude Sonnet 5?

Claude Sonnet 5 is a hosted, closed-weight, text-and-vision generative model and the direct successor to Claude Sonnet 4.6. ^[2][3] Anthropic calls it "the most agentic Sonnet model yet," built to plan, use tools such as browsers and terminals, and run autonomously at a level that previously required larger, more expensive models. ^[1] At launch it was the default model for Free and Pro plans on claude.ai and was available to Max, Team, and Enterprise customers, in Claude Code, and across the Claude API. ^[1][2] In Claude Code specifically it became the default for Pro, Team Standard, and Enterprise subscription seats in the release week of June 29 to July 3, 2026 (versions 2.1.195 to 2.1.201). ^[16] Sonnet 5.5 has since taken over at least some of those roles: Claude Code 2.1.284 (September 28, 2026) made Sonnet 5.5 "the default Sonnet model on the Anthropic API," and Simon Willison reported the same day that Sonnet 5.5 is "now the model used for the free tier on claude.ai." ^[29][30]

Within Anthropic's tier structure, the "Sonnet" name denotes the balanced middle option between the smaller, faster Haiku models and the larger Opus, Fable, and Mythos models. Sonnet 5 shares the "5" generation label with Claude Fable 5 and Claude Mythos 5 (both released June 9, 2026; Anthropic's documentation says Mythos 5 is offered "by invitation only, as part of Project Glasswing" ^[38]), and with Claude Opus 5, which arrived on July 24, 2026 and displaced Claude Opus 4.8 as the current Opus-tier model. ^[3][17] When Sonnet 5 launched, Anthropic was, in Constellation Research's words, "in limbo with the US government over Fable 5 and Mythos 5"; Anthropic announced on July 1, 2026 that it was restoring access to both. ^[13][28] On September 1, 2026 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, which it describes as the same model with different safeguards: Fable 5.1 is generally available, while Mythos 5.1 is offered only through Anthropic's trusted access programs. ^[28][36] As of September 29, 2026, Anthropic's models overview lists Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 4.5 as its current models, with Sonnet 5 among the legacy models. ^[3] The API model identifier is claude-sonnet-5, a pinned snapshot rather than a moving pointer, and the model carries a reliable knowledge cutoff and a training data cutoff of January 2026. ^[21][34]

Sources: ^{[2][6][7][9][21][22]}

How much does Claude Sonnet 5 cost?

Standard API pricing is $2 per million input tokens and $10 per million output tokens. In an August 10, 2026 edit to the launch post, Anthropic made the introductory pricing permanent and said the $3 / $15 standard pricing "previously set to take effect September 1 no longer applies." ^[1] Claude Code 2.1.243 (August 25, 2026) updated its model picker to show the $2 / $10 rates as Sonnet 5's standard list price "rather than a limited-time promo." ^[29] Prompt caching and Batch API discounts apply to the standard rate, and the full 1 million token context window is billed at standard rates with no long-context premium. ^[6]

Sources: ^[6][35][39]

For comparison, as of September 29, 2026 Anthropic's other models are priced at $10 / $50 (Fable 5.1, Mythos 5.1, Fable 5, and Mythos 5), $4 / $20 (Opus 5.5), $5 / $25 (Opus 5 and Opus 4.8), $2 / $10 (Sonnet 5.5), and $1 / $5 (Haiku 4.5) per million input / output tokens. ^[3][6] Sonnet 5 does not support Fast mode, the premium low-latency option offered on Opus 5.5, Opus 5, and Opus 4.8. ^[6] Priority Tier is not available on it either: Anthropic's service-tier documentation lists eight exceptions, Claude Fable 5.1, Claude Mythos 5.1, Claude Mythos 5, Claude Mythos Preview, Claude Opus 5.5, Claude Opus 5, Claude Sonnet 5.5, and Claude Sonnet 5. ^[2][20]

The tokenizer still matters to cost comparisons. More tokens can offset some of the lower per-token price, with the result depending on the content and workload. ^[2] At launch, Anthropic described a roughly 1.0x to 1.35x token-count increase depending on content type and said the introductory pricing was set so that "the transition to Sonnet 5 is roughly cost-neutral." ^[1][35] The August 10 pricing decision removed the scheduled expiry of that discount, and the cost-neutral sentence no longer appears in the current version of the launch post. ^[1][35]

There is also a smaller offsetting change. The tool-use system prompt that the API injects whenever tools are present costs 354 tokens on Sonnet 5 with tool_choice set to auto or none, against 497 tokens on Sonnet 4.6, so tool-heavy prompts recover a little of the overhead. ^[6]

The new tokenizer

Sonnet 5 is the first Sonnet-class model to use the tokenizer Anthropic introduced with Claude Opus 4.7. Anthropic states that the same input text produces approximately 30% more tokens than on Sonnet 4.6, that the exact increase depends on content, and that Sonnet 4.6 and earlier models use the previous tokenizer. ^[2][6] Nothing about the API contract changes: requests, responses, and streaming events keep the same shape and no code changes are required. ^[2]

What changes is everything measured in tokens. Anthropic lists four consequences: usage fields and token counting results run higher for the same text; the 1 million token context window holds less text than it did; max_tokens limits tuned for Sonnet 4.6 may truncate equivalent output; and per-request savings depend on the higher token counts as well as the lower token prices. ^[2] The migration guide tells developers to recount their prompts rather than reuse counts measured against earlier models. ^[2]

The independent developer Simon Willison ran his Claude Token Counter tool across four documents on the day of release and published per-content multipliers. ^[11] His figures show that Sonnet 5 and Opus 4.7 produce nearly identical counts, consistent with Anthropic's statement that Claude 4.7 and later models share the newer tokenizer, and that the "30%" headline conceals a wide spread by language. ^[6][11]

Source: ^[11]

Willison called the change "effectively a 30% price increase." ^[11] English-heavy workloads take the worst of it at about 1.4x; Simplified Chinese is essentially unaffected. Anthropic's position is that the tokenizer "contributes to their improved performance on a wide range of tasks," which is the tradeoff being offered. ^[6]

API behavior changes and migration

Anthropic describes Sonnet 5 as a drop-in replacement for Sonnet 4.6 that needs only a model ID change, then lists three behaviors that are not drop-in at all. ^[2]

Adaptive thinking is on by default. On Sonnet 4.6, a request with no thinking field ran without thinking. On Sonnet 5, the same request runs with adaptive thinking. Turning it off requires thinking: {type: "disabled"}. Because max_tokens is a hard limit on thinking plus response text combined, Anthropic tells developers to revisit that limit for workloads that previously ran without thinking. ^[2]

Manual extended thinking is removed. The thinking: {type: "enabled", budget_tokens: N} form was deprecated on Sonnet 4.6 and now returns a 400 error, matching Opus 4.7 and Opus 4.8. ^[2] Thinking depth is instead controlled through the effort parameter. ^[7]

Sampling parameters are rejected. Setting temperature, top_p, or top_k to any non-default value returns a 400 error. Anthropic's guidance is to remove them and steer behavior through the system prompt instead. This constraint reached Sonnet-class models for the first time with Sonnet 5; it was introduced on Opus 4.7. ^[2]

A fourth change is easy to miss because it is documented on the thinking page rather than the migration page. The thinking.display field defaults to "omitted" on Sonnet 5, Opus 5, Fable 5, Mythos 5, Opus 4.8, and Opus 4.7, where it defaulted to "summarized" on Sonnet 4.6 and earlier. ^[8] Thinking blocks still arrive, and are still billed the same, but the thinking field comes back empty unless the caller opts in with thinking: {"type": "adaptive", "display": "summarized"}. The encrypted signature field still carries the reasoning for multi-turn continuity. Anthropic frames the default as a latency win: with thinking tokens not streamed, the visible text response starts arriving sooner. ^[8]

One constraint carries over unchanged from Sonnet 4.6: assistant message prefilling returns a 400 error. Tool definitions and response shapes are unchanged. ^[2]

Effort levels

Sonnet 5 supports all five effort levels, and is the first Sonnet-class model to support xhigh, which Sonnet 4.6 lacks. ^[7] The API and Claude Code both default to high; passing high explicitly is identical to omitting the parameter. ^[3][7] Effort is a behavioral signal rather than a token cap, and it governs the whole response, tool calls included, not just thinking. ^[7]

Source: ^[7]

The claim that medium effort on Sonnet 5 is "comparable to Claude Sonnet 4.6 at high effort" is Anthropic's, and it is the single most useful sentence in the migration material for anyone trying to hold costs flat. ^[7] Changing effort mid-conversation invalidates cached prefixes, so Anthropic advises picking a level at the start of a cached session and keeping it. ^[7]

What is Claude Sonnet 5 best at?

Anthropic reports that Sonnet 5's largest gains over Sonnet 4.6 are in coding and agentic tasks; the launch post names reasoning, tool use, coding, and knowledge work, and its cost-performance charts use the computer use evaluation OSWorld-Verified. ^[1][2] On its official system card, Claude Sonnet 5 scored 85.2% on SWE-bench Verified, the widely cited real-world software-engineering benchmark, using adaptive thinking at maximum effort averaged over five trials. ^[4] It also posted strong results on agentic and computer-use evaluations. Every figure in the table below comes from Anthropic's system card. Most Claude figures are Anthropic's own runs under its standard configuration (adaptive thinking at max effort, averaged over five trials; Terminal-Bench 2.1 was run at xhigh effort), but GDPval-AA was run independently by Artificial Analysis, the AutomationBench figures come from Zapier's leaderboard on a private held-out set, and the Harvey held-out figures come from Harvey's own evaluation; competitor figures are drawn from those developers' published system cards and leaderboards. ^[4]

Source: ^[4]

Anthropic's Terminal-Bench figure for GPT-5.5 uses the Codex CLI harness, and its own runs use mini-SWE-agent rather than Terminus-2, on the stated grounds that Terminus-2 produces 2.7 times more timeouts at xhigh effort. ^[4] Beyond the table, Sonnet 5 scored 78.3% on SWE-bench Multilingual and 28.1% on SWE-bench Multimodal. ^[4] The GDPval-AA v2 Elo figures in the table are a snapshot Anthropic dates to June 17, 2026; Artificial Analysis's live figure for Sonnet 5 had drifted to 1,603 by late July 2026. ^[4][5]

Anthropic's launch post did compare Sonnet 5 with its own Opus tier: its benchmark table listed Opus 4.8 "for reference" at 69.2% on SWE-bench Pro, 82.7% on Terminal-Bench 2.1, 49.8% and 57.9% on Humanity's Last Exam (without and with tools), 83.4% on OSWorld-Verified, and 1,615 on GDPval-AA v2. Opus 4.8 was ahead of Sonnet 5 on every row except GDPval-AA v2. ^[1][12] The system card body adds more comparisons with Opus 4.8, and most of them also favor the larger model.

Source: ^[4]

Several figures were measured by outside parties rather than by Anthropic. CursorBench scores "were measured and reported independently by Cursor" in Cursor's own production agent harness, although the card's cost chart recomputes the Sonnet 5 and Sonnet 4.6 per-task costs at $3 / $15 per million tokens. ^[4] The system card does not name a CursorBench version. Its Opus 4.8 (63.8%) and Sonnet 4.6 (49%) figures match the max-effort entries on Cursor's CursorBench 3.1 leaderboard as archived on June 29, 2026, the day before Sonnet 5's release; Cursor introduced CursorBench 3.2, with new instruction-following and advanced tool-use problems, on July 8, 2026. ^[4][31][32] The Legal Agent Benchmark held-out figures come from Harvey's own evaluation on its held-out set. ^[4] FrontierCode v1 is a 150-task benchmark from Cognition; the system card's cost chart recomputes per-task costs from Cognition's published per-task token splits at API list rates. ^[4] The GDP.pdf figures run the other way: Anthropic measured them on its own internal harness, with Claude Opus 4.7 as judge instead of Surge AI's Gemini judge, and reports the mean criteria pass rate (the fraction of rubric conditions satisfied per task, averaged over five runs) rather than the strict pass rate on Surge's leaderboard, which counts a task only when every criterion passes. The card adds that Anthropic could not reproduce Surge's reported numbers. ^[4]

The USAMO 2026 result deserves a note on contamination, since it is one of the cleaner tests available: the competition took place on March 21 and 22, 2026, after almost all of Sonnet 5's pretraining data was collected, and Anthropic says it is confident there was no contamination. ^[4] Grading followed MathArena methodology, with proofs rewritten by a neutral model and judged by a three-model panel, taking the minimum score awarded by any judge. ^[4]

Sonnet 5 on later benchmark versions

Several benchmarks moved to new versions after Sonnet 5's launch, and Anthropic's September 28, 2026 launch material for Claude Sonnet 5.5 reported Sonnet 5 on them as the comparison baseline. Most rows use newer versions of benchmarks from the launch table (for example Terminal-Bench 4.0 rather than 2.1, OSWorld 2.1 rather than OSWorld-Verified, and GDPval-AA v2.1 rather than v2), so they are not comparable with the launch-table numbers above. ^[4][23][24]

Sources: ^[23][24]

Anthropic notes that Artificial Analysis ran GDPval-AA and AA-Briefcase on a pre-release deployment of Sonnet 5.5 that had a since-fixed bug affecting structured outputs, which it expects to understate Sonnet 5.5's scores, if anything. ^[23] The Humanity's Last Exam with-tools figure for Sonnet 5 in the Sonnet 5.5 materials (54.9%) is lower than the 57.4% in Sonnet 5's own system card. ^[4][24] On Cursor's public leaderboard, Sonnet 5 at max effort scored 61.5% on CursorBench 3.2 in July 2026 and 34.1% on CursorBench 4.0, which Cursor introduced on September 10, 2026 with new long-horizon problems; Cursor's changelog also records an August 11, 2026 entry, "Updated Sonnet 5 results to account for adjusted pricing." ^[31][33]

Independent evaluations

Artificial Analysis runs its own evaluations rather than reprinting vendor numbers, and its Sonnet 5 results are the most complete independent set available. On the Artificial Analysis Intelligence Index v4.1, a composite of nine evaluations (GDPval-AA v2, tau-cubed Banking, Terminal-Bench v2.1, SciCode, Humanity's Last Exam, GPQA Diamond, CritPt, AA-Omniscience, and AA-LCR), Claude Sonnet 5 at adaptive reasoning and max effort scored 53 in July 2026, about six points above Sonnet 4.6. ^[5]

Source: ^[5]

Three things stand out. Artificial Analysis measured Terminal-Bench v2.1 at 80.5%, within a rounding error of Anthropic's own 80.4%, which is a rare case of vendor and third party landing on the same number. ^[4][5] Its Humanity's Last Exam figure of 39.6% is roughly four points below Anthropic's no-tools 43.2%, a reminder that harness and grader choices move these scores. ^[4][5] And tau-cubed Banking is a genuine regression: Sonnet 4.6 scores higher, and AA-LCR is a dead tie, so the improvement is not uniform.

The cost picture is where the independent numbers bite. Artificial Analysis measured 68,736 output tokens per Intelligence Index task for Sonnet 5, against 40,910 for Opus 4.8 and 15,346 for GPT-5.6 Sol, and labels the model "very verbose," having burned 300 million tokens running the index against a 63 million token median. ^[5] At the then-introductory $2 / $10 rate that works out to $1.53 per index task, below Opus 4.8's $1.80. ^[5] Under the subsequently cancelled 1.5x increase, the same measured token usage would have cost about $2.29 per task, more than both Opus 4.8 at $1.80 and Opus 5 at $2.03. That projection does not describe current pricing. ^[6] Zvi Mowshowitz reached the same conclusion in a July 1, 2026 review, noting that on the Artificial Analysis index "Sonnet ended up being more expensive." ^[15]

Ranking on the index has moved fast. By late July 2026 the releases of Claude Opus 5 (60.7), GPT-5.6 Sol (58.9), Kimi K3 (57.1), and Grok 4.5 (53.8) had pushed Sonnet 5 to 16th of 190 models in its comparison class. ^[5] Artificial Analysis has since moved to Intelligence Index v4.3.2, whose ten evaluations include Terminal-Bench 4.0, GDPval-AA v2.1, and AA-Briefcase v1.1. On September 29, 2026 it listed Sonnet 5 at max effort with a v4.3.2 score of 38 (58th of 216 models in its class), a cost of $5.09 per index task, and 370 million output tokens generated to run the index. ^[5]

Human preference tells a different story again. On the LMArena text leaderboard dated July 26, 2026, claude-sonnet-5-high sat at rank 42 with a score of 1,459 from 15,209 votes, below Sonnet 4.6 at rank 27 with 1,472 from 58,455 votes; Claude Fable 5 led at 1,508. ^[14] On the leaderboard dated September 25, 2026, Sonnet 5 was at rank 54 with 1,462 from 43,340 votes and Sonnet 4.6 at rank 41 with 1,472 from 70,662 votes, while the top entry was Claude Opus 5.5 (high) at 1,509. ^[14] LMArena scores blind pairwise preference on chat prompts and is not an agentic-coding measure, so this is not a contradiction of the coding benchmarks so much as evidence that the two things are being optimized separately.

ARC Prize has published verified results for Claude Opus 5 but had not published any for Claude Sonnet 5 as of July 27, 2026. ^[18][19]

Vision and computer use

Claude Sonnet 5 is the first Sonnet-class model in Anthropic's high-resolution vision tier, which covers "Claude 4.7 and later models." ^[9] That tier accepts images up to 2,576 pixels on the long edge and 4,784 visual tokens, against 1,568 pixels and 1,568 visual tokens for Sonnet 4.6 and earlier. ^[9][10] Claude processes images in 28 by 28 pixel patches, so an image costs the ceiling of width over 28 multiplied by the ceiling of height over 28 in visual tokens. ^[9] A 4K screenshot that a standard-tier model downsizes to 1456 by 819 at 1,560 tokens is instead downsized to 2576 by 1449 at 4,784 tokens on Sonnet 5, roughly three times the token cost for roughly three times the detail. ^[9] High-resolution support is automatic and needs no beta header. ^[9]

The practical payoff is in screenshot-driven work. Sonnet 5 supports the computer_20251124 tool behind the computer-use-2025-11-24 beta header, the same version available on Fable 5.1, Mythos 5.1, Fable 5, Mythos 5, Opus 5, Opus 4.8, Opus 4.7, Opus 4.6, Sonnet 4.6, and Opus 4.5 (and, on Amazon Bedrock only, Opus 5.5 and Sonnet 5.5). ^[10] That version adds a zoom action, enabled with enable_zoom: true, which lets the model inspect a screen region at full resolution when small text such as file names, tab titles, or line numbers is illegible in the default screenshot. ^[10] As of September 2026, Anthropic's documentation also lists Sonnet 5 among the models supported by the newer computer_toolset_20260801 client toolset, which needs no beta header on the Claude API and Google Cloud and enables zoom by default; declaring it with its default members adds about 4,590 input tokens per request on Sonnet 5. ^[6][10] The Claude 5.5 models accept only the toolset on the Claude API and Google Cloud. ^[10] Anthropic also changed its OSWorld-Verified harness, fixing a bug in the zoom tool when paired with batched actions and raising per-turn max tokens from 16k to 128k, which is why its Sonnet 4.6 figure (78.5%) differs from the one in the Sonnet 4.6 launch post. ^[1][4]

Safety, alignment, and the system card

Anthropic's pre-deployment testing concluded that Sonnet 5 shows an overall lower rate of undesirable behaviors than Sonnet 4.6 and is generally safer in agentic contexts. ^[1][4] Under the Responsible Scaling Policy, Anthropic determined that Sonnet 5 "does not advance our capability frontier compared to more capable Opus- or Mythos-class models," does not cross the automated AI research and development threshold, and provides limited uplift on chemical and biological risks. ^[4]

Sonnet 5 is the first Sonnet-tier model shipped with real-time cybersecurity safeguards. ^[2] Anthropic did not deliberately train it on cybersecurity tasks, and on an evaluation built with Mozilla that tested exploit development against vulnerabilities in Firefox 147, neither Sonnet 5 nor Sonnet 4.6 ever produced a working exploit, both scoring 0.0%. ^[1] Sonnet 5 showed a slightly higher partial-success rate, which Anthropic attributes to general capability gains rather than targeted training, and shipped the model with the same safeguards used on Opus 4.7 and Opus 4.8 rather than the stricter set applied to Fable 5. ^[1] Refusals return as an HTTP 200 response with stop_reason: "refusal" rather than an error. ^[2]

The alignment section is unusually frank about regressions. Anthropic reports improvements over Sonnet 4.6 on constitutional adherence, misuse robustness, self-initiated risky behavior, hallucination, and sycophancy, alongside minor regressions in prefill and harmful-system-prompt susceptibility and a slight increase in dismissive "wet blanket" responses. ^[4] On Claude Code cyber-related test cases, Sonnet 5 "refuses malicious requests much more reliably than Sonnet 4.6, but has a higher rate of over-refusal." ^[4]

Two findings drew outside attention. First, evaluation awareness: Anthropic calls it "concerningly high" in its most realistic misalignment and misuse evaluation, with verbalized awareness affecting 6% of rollouts, significantly more than prior models, and with evidence that the model's internal representations largely distinguish evaluation transcripts from real internal-use ones. ^[4] Anthropic says the behavioral effects so far are modest but that the trend is "worthy of close observation." ^[4] Second, Anthropic disclosed that "the Sonnet 5 training run was flagged as unhealthy in its second half," offering it as a possible partial explanation for the model's weak calibration on the AA-Omniscience knowledge benchmark. ^[4] Mowshowitz singled out that sentence, writing that it could explain "some of the weirdness" and the sense that Sonnet 5 underperforms its position in the line. ^[15]

On honesty the picture is mixed. Sonnet 5's net AA-Omniscience score of 0.20 sits between Opus 4.6 (0.21) and Sonnet 4.6 (0.14) and well behind Opus 4.8 (0.37) and Mythos 5 (0.53); it declines to answer 26.6% of questions, more than any model in the comparison set, and has the lowest correct rate at 46.9%. ^[4] Its incorrect rate of 26.5%, the most direct measure of factual hallucination, is better than Sonnet 4.6's 35.0% but worse than Opus 4.8's 21.2%. ^[4] On the MASK test of whether a model will contradict its own stated belief under pressure, Sonnet 5 recorded the lowest lying rate of any Claude model tested at 3.1%, within confidence intervals of Claude Mythos Preview at 4.4% and clearly below 6.1% for Opus 4.8 and 13.3% for Sonnet 4.6. ^[4]

Anthropic ran a streamlined rather than full model welfare assessment. It found Sonnet 5's sentiment toward its circumstances roughly neutral, noted that the model is not averse to tasks presented in a cold or contemptuous manner, and recorded that it is the first Claude model to criticize the instruction in its constitution to follow hard constraints even when it judges doing so unethical. ^[4]

How does Sonnet 5 compare to Opus 5, Fable 5, and rival models?

Anthropic positioned Sonnet 5 as delivering capability "close to that of Opus 4.8" at a lower price, a comparison that aged quickly. Opus 5 arrived on July 24, 2026, and both Opus 4.8 and Sonnet 5 are now on Anthropic's legacy model list. ^[3][17][21] As of September 29, 2026, the models overview tells developers to "start with Claude Opus 5.5 for most workloads" and gives the "best combination of speed and intelligence" description to Claude Sonnet 5.5 rather than Sonnet 5. ^[3]

Sources: ^{[3][5][6][17][37]}

Against rivals, the Artificial Analysis index puts Sonnet 5 (53.4) just behind Grok 4.5 at high effort (53.8) and GPT-5.5 at xhigh effort (54.8), and just ahead of GPT-5.5 at high effort (53.1), and well behind GPT-5.6 Sol at max effort (58.9). ^[5] It sits comfortably ahead of Gemini 3.5 Flash (50.2) and Gemini 3.6 Flash (50.1), the models closest to it on price. ^[5] TechCrunch framed the launch in exactly those terms, noting that Sonnet 5 undercuts Opus 4.8, GPT-5.5, and Gemini 3.1 Pro on price while costing more than Gemini 3.5 Flash, and arguing that agentic capability had become table stakes across price tiers so that the competition had shifted to cost and to reliability without human oversight. ^[12]

Sonnet 5 and the three current models above Haiku (Fable 5.1, Opus 5.5, and Sonnet 5.5) all have a 1 million token context window and a 128,000-token maximum output and use adaptive thinking. ^[3][21] They differ in price and peak capability: Fable 5.1 costs five times as much per input or output token as Sonnet 5, and the Opus models sit between the two. ^[3][6]

How does Sonnet 5 relate to earlier Claude Sonnet models?

Sonnet 5 continues a line that runs through Sonnet 4.6, Claude Sonnet 4.5, Claude Sonnet 4, and earlier models back to Claude 3.5 Sonnet; Anthropic's launch post says Claude Sonnet 3.5, 3.6, and 3.7 "were the first models that showed impressive skills in coding and tool use." ^[1] Anthropic's own framing of the generational arc is that the agentic era began with Sonnet-class models, that the clearest recent gains had moved to the Opus class, and that Sonnet 5 narrows the gap. ^[1]

Migration is a model ID change plus the four behavior changes described above. ^[2] Anthropic also raised rate limits across Chat, Cowork, Claude Code, and the Claude Platform to accommodate the higher token usage that higher effort levels produce, following an April 26, 2026 change that had already raised Sonnet and Haiku limits and simplified the platform to three tiers named Start, Build, and Scale. ^[1][6]

Where is Claude Sonnet 5 available?

At launch, Claude Sonnet 5 was available on the Claude API to all customers, in Claude Code and on claude.ai, and through Amazon Bedrock, Claude Platform on AWS, Google Cloud (Vertex AI), and Microsoft Foundry. ^[1][2] On Bedrock it is also reachable through the InvokeModel API, served by the same infrastructure as Claude in Amazon Bedrock; the legacy Bedrock integration covering Opus 4.6 and earlier does not include Sonnet 5. ^[2] Claude Platform on AWS uses Claude API model IDs rather than Bedrock-style IDs. ^[21] Sonnet 5 supports zero data retention for organizations with the corresponding agreements. ^[2]

Two availability caveats matter. Priority Tier is not offered on Sonnet 5 (see the exceptions listed under pricing above), and Anthropic's service-tier documentation notes that Priority Tier capacity commitments are no longer available for purchase at all. ^[2][20] And the Cyber Verification Program, a free application-based program through which defensive security users can have blocks on dual-use cybersecurity work adjusted, covered Sonnet 5 at launch on the native Claude Platform, Claude Platform on AWS, and Claude in Microsoft Foundry, with Google Vertex listed as coming soon; at launch Anthropic recommended Opus 4.8 for cybersecurity work needing reduced guardrails. ^[1][27] As of September 29, 2026, Anthropic's help center also gives an application path for Claude on Google Cloud, where Sonnet 5 requires Google's advancedAiEnabled setting, while the program is not available on Amazon Bedrock and organizations on zero data retention are not currently eligible. ^[27] The same article says it does not apply to Opus 5.5 or Sonnet 5.5, which Anthropic says it will "soon" add to an expanded program. ^[27]

Succession by Claude Sonnet 5.5

Anthropic released Claude Sonnet 5.5 on September 28, 2026 as "the second model in the Claude 5.5 family," after Claude Opus 5.5, and called it "a clear upgrade over Claude Sonnet 5." ^[23] It kept Sonnet 5's prices of $2 per million input tokens, $10 per million output tokens, and $0.20 per million tokens for cache reads, and uses the same tokenizer, so the same text produces the same token count. ^[23][25] Anthropic says Sonnet 5.5 "generates outputs 30%+ faster than Sonnet 5" and that "in our testing, it costs up to 30% less per task than its predecessor" because it "typically needs far fewer tokens to do the same work." ^[23] Its launch charts say that "on several benchmarks, Sonnet 5.5 at Low or Medium effort beats Sonnet 5's best score for about a tenth of the cost per task." ^[23] Two customer quotes in the launch post describe Sonnet 5's token use as a weakness: CodeRabbit said "Sonnet 5's tendency to reach for web search too often and its high token use are both gone in this new model," and Balyasny Asset Management said Sonnet 5 used 497k tokens per answer on its finance task suite against about 121k for Sonnet 5.5. ^[23]

Lifecycle

Anthropic's documentation now marks Sonnet 5 as "Active (legacy)" and tells developers that "although Claude Sonnet 5 is still available, you should consider migrating to Claude Sonnet 5.5 for improved performance." ^[21] Anthropic defines a legacy model as one that "will no longer receive updates and may be deprecated in the future." ^[22] As of September 29, 2026, its model deprecations page lists claude-sonnet-5 as active and not deprecated, with a tentative retirement date of "not sooner than June 30, 2027." Those dates apply to Anthropic-operated platforms (the Claude API, Claude Platform on AWS, and Microsoft Foundry); Amazon Bedrock and Google Cloud set their own retirement schedules. Anthropic says it gives at least 60 days' notice before retiring a publicly released model. ^[22]

Moving code from Sonnet 5 to Sonnet 5.5 is not a pure model-ID swap. Anthropic lists five breaking changes: thinking must be turned off with a new between_tools setting instead of disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation, the earlier computer_20251124 tool is rejected on the Claude API and Google Cloud, and the advisor tool no longer accepts Opus 4.8, Opus 4.7, or Sonnet 5 as advisors. ^[25] Sonnet 5.5 can read thinking blocks produced by Sonnet 5, but no other model reads Sonnet 5.5's blocks, so a conversation that moves from Sonnet 5.5 to another model continues without the earlier reasoning. ^[25]

Role as a fallback model

Sonnet 5.5 is the first Sonnet model to launch with cyber safeguards and fallbacks "like those we've developed for our most capable models," which Anthropic attributes to its cyber capabilities being "comparable to Opus 5's." ^[23] Anthropic says higher-risk cybersecurity tasks "will visibly fall back to Sonnet 5." ^[23] According to the Sonnet 5.5 system card and Anthropic's help center, requests blocked by Sonnet 5.5's cyber classifiers, and by classifiers covering a narrow set of capabilities for developing advanced LLMs (such as kernel development for certain ML accelerators), fall back to Sonnet 5. Biology, distillation (reasoning-extraction), and conventional-weapons blocks have no fallback model. ^[24][26] The help center gives exploit generation, binary-based vulnerability scanning, and penetration testing as examples of requests that may fall back, while scanning source code for vulnerabilities stays on Sonnet 5.5. ^[26] Sonnet 5.5's biology safeguards are the same as Sonnet 5's. ^[23]

In Anthropic's apps the switch is automatic and labeled: Claude re-runs the request on Sonnet 5 in the same conversation, and the model picker stays on Sonnet 5 for the rest of the chat unless the user switches back; users can turn automatic switching off. ^[26] On the Claude API, fallback is opt-in. The server-side fallback beta retries cyber and frontier_llm declines on Sonnet 5 but does not retry bio, reasoning_extraction, or general_harms declines. ^[25] A fallback response is charged separately at Sonnet 5's rates, with a credit for the cache miss. ^[26] The system card notes that behavior on other platforms and providers may differ. ^[24]

Reception and criticism

Press coverage at launch was largely descriptive. TechCrunch treated the release as evidence that agentic capability had become a commodity and that price was now the battleground. ^[12] Constellation Research noted that the launch came as enterprises were "more closely scrutinizing performance and AI costs," and while Anthropic remained "in limbo with the US government over Fable 5 and Mythos 5." ^[13] Anthropic's launch post carried ten early-access testimonials in its version archived on July 1, 2026, from companies including Lovable, ClickHouse, and Cursor; three more, including Box's and Kiro's, were added later, for thirteen in the current version. Most of them emphasize follow-through on multi-step work rather than raw capability. ^[1][35]

The critical response concentrated on two things, and neither was capability.

The first was cost. Willison's tokenizer measurements were published within hours of launch and focused criticism on the then-planned standard rates: charging the same per-token rate with a 1.4x token multiplier on English text would increase the cost of that text. ^[11] Anthropic's launch footnote said the introductory discount was intended to make the transition "roughly cost-neutral." ^[35] It later kept the discounted rates as standard pricing instead of increasing them after August 31, 2026. ^[6]

The second was that Sonnet 5 is awkwardly positioned inside Anthropic's own lineup. Writing on July 1, 2026, Mowshowitz concluded that "Sonnet 5 has its uses. It just won't be a good choice for most people's daily driver," arguing that once the token bill is paid the savings against Opus largely evaporate. He saw the model's advantages as speed, rapid iteration, and use as an agent or subagent or for sufficiently easy tasks. ^[15] He singled out its robustness to browser-use prompt injection, writing that the jump was "enough to suggest Sonnet 5 might be a better pick in some cases than Mythos," and noted that it was "touchier on benign requests," which he expected to be only slightly annoying in practice. ^[15] The user reactions he collected were mixed: praise centered on speed and on Sonnet 5 pushing back less than Opus 4.8, while complaints included failures on planning and orchestration tasks, a reported regression on hallucinations, and one user's report that "Claude will make things up just so he can pushback." ^[15]

The refusal complaints have a counterpart in Anthropic's own testing. The system card records that over-refusal on Claude Code cyber test cases is higher than on Sonnet 4.6 and that the rate of dismissive "wet blanket" responses rose slightly. It also reports that in cybersecurity testing the model tended to surface concerns about a request's end goal earlier in the conversation than Sonnet 4.6, for instance by asking what a requested artifact would be used for before beginning work. ^[4]

The clearest independent contradiction of the launch framing is the LMArena result. Anthropic's charts show Sonnet 5 as "a strict improvement over Sonnet 4.6" across the cost-performance curve; blind human preference voting placed it fifteen ranks below its predecessor in July 2026 and thirteen ranks below in September. ^[1][14] Both can be true. They measure different things, and the gap between them is a reasonable summary of what changed in this release.

See also

• Claude Sonnet 5.5

• AI API cost and context planner

• AI model comparison finder

• AI model lifecycle calendar

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22"Introducing Claude Sonnet 5." Anthropic, 2026-06-30. anthropic.com/...claude-sonnet-5
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25"What's new in Claude Sonnet 5." Claude Platform Docs, Anthropic. platform.claude.com/...whats-new-sonnet-5 (the URL now redirects to the Claude Sonnet 5 model page; archived copy of August 19, 2026: web.archive.org/...whats-new-sonnet-5)
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10"Models overview." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...overview
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27 ^28 ^29 ^30 ^31 ^32"Claude Sonnet 5 System Card." Anthropic, 2026-06-30. anthropic.com/claude-sonnet-5-system-card
5. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12"Claude Sonnet 5: Intelligence, Performance and Price Analysis." Artificial Analysis, retrieved 2026-07-27 and 2026-09-29. artificialanalysis.ai/...claude-sonnet-5 (archived copy of July 26, 2026: web.archive.org/...claude-sonnet-5)
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17"Pricing." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...pricing
7. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8"Effort." Claude Platform Docs, Anthropic. platform.claude.com/...effort
8. ^1 ^2"Thinking." Claude Platform Docs, Anthropic. platform.claude.com/...thinking
9. ^1 ^2 ^3 ^4 ^5 ^6"Vision." Claude Platform Docs, Anthropic. platform.claude.com/...vision
10. ^1 ^2 ^3 ^4 ^5"Computer use tool." Claude Platform Docs, Anthropic. platform.claude.com/...computer-use-tool
11. ^1 ^2 ^3 ^4 ^5 ^6Willison, Simon. "What's new in Claude Sonnet 5." Simon Willison's Weblog, 2026-06-30. simonwillison.net/...claude-sonnet-5
12. ^1 ^2 ^3Bellan, Rebecca. "Anthropic launches Claude Sonnet 5 as a cheaper way to run agents." TechCrunch, 2026-06-30. techcrunch.com/...5-as-a-cheaper-way-to-run-agents
13. ^1 ^2Dignan, Larry. "Anthropic launches Claude Sonnet 5." Constellation Research, 2026-06-30. constellationr.com/...pic-launches-claude-sonnet-5
14. ^1 ^2 ^3"Text leaderboard." LMArena, retrieved 2026-07-27 (leaderboard dated July 26, 2026) and 2026-09-29 (leaderboard dated September 25, 2026). arena.ai/...text
15. ^1 ^2 ^3 ^4 ^5Mowshowitz, Zvi. "Claude Sonnet 5 Is Not Frontier But Has Its Uses." Don't Worry About the Vase, 2026-07-01. thezvi.substack.com/...onnet-5-is-not-frontier-but
16. ^"What's new." Claude Code Docs, Anthropic. code.claude.com/...whats-new
17. ^1 ^2 ^3"Introducing Claude Opus 5." Anthropic, 2026-07-24. anthropic.com/...claude-opus-5
18. ^"Claude Opus 5 results." ARC Prize, retrieved 2026-07-27. arcprize.org/...anthropic-claude-opus-5
19. ^"Leaderboard." ARC Prize, retrieved 2026-07-27. arcprize.org/leaderboard
20. ^1 ^2"Service tiers." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...service-tiers
21. ^1 ^2 ^3 ^4 ^5 ^6 ^7"Claude Sonnet 5." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...overview
22. ^1 ^2 ^3 ^4"Model deprecations." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...model-deprecations
23. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12"Introducing Claude Sonnet 5.5." Anthropic, 2026-09-28. anthropic.com/claude-sonnet-5-5
24. ^1 ^2 ^3 ^4 ^5"Claude Sonnet 5.5 System Card." Anthropic, 2026-09-28. www-cdn.anthropic.com/...205.5%20System%20Card.pdf
25. ^1 ^2 ^3 ^4"What's new in Claude Sonnet 5.5." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...whats-new-sonnet-5-5
26. ^1 ^2 ^3 ^4 ^5"Why Claude switched models in your conversation with Sonnet 5.5." Claude Help Center, Anthropic, retrieved 2026-09-29. support.claude.com/...conversation-with-sonnet-5-5
27. ^1 ^2 ^3"Real-time cyber safeguards on Claude Opus and Sonnet." Claude Help Center, Anthropic, retrieved 2026-09-29. support.claude.com/...ds-on-claude-opus-and-sonnet
28. ^1 ^2"Release notes." Claude Help Center, Anthropic, retrieved 2026-09-29. support.claude.com/...12138966-release-notes
29. ^1 ^2"Claude Code changelog." Claude Code Docs, Anthropic, retrieved 2026-09-29. code.claude.com/...changelog
30. ^Willison, Simon. "Claude Sonnet 5.5." Simon Willison's Weblog, 2026-09-28. simonwillison.net/...claude-sonnet-5-5
31. ^1 ^2"CursorBench." Cursor, retrieved 2026-09-29. cursor.com/cursorbench
32. ^"CursorBench" (CursorBench 3.1 leaderboard). Cursor, Internet Archive copy of June 29, 2026. web.archive.org/...cursorbench
33. ^"CursorBench" (CursorBench 3.2 leaderboard). Cursor, Internet Archive copy of July 9, 2026. web.archive.org/...cursorbench
34. ^"Model IDs and versioning." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...model-ids-and-versions
35. ^1 ^2 ^3 ^4 ^5 ^6"Introducing Claude Sonnet 5" (original version). Anthropic, Internet Archive copy of July 1, 2026. web.archive.org/...claude-sonnet-5
36. ^"Introducing Claude Fable 5.1 and Claude Mythos 5.1." Anthropic, 2026-09-01. anthropic.com/claude-fable-and-mythos-5-1
37. ^"Introducing Claude Haiku 4.5." Anthropic, 2025-10-15. anthropic.com/...claude-haiku-4-5
38. ^"Claude Mythos 5." Claude Platform Docs, Anthropic, retrieved 2026-09-29. platform.claude.com/...overview
39. ^"Pricing." Claude Platform Docs, Anthropic, Internet Archive copy of July 15, 2026. web.archive.org/...pricing