Claude Opus 4.8

Claude Opus 4.8 is a large language model developed by Anthropic, released on May 28, 2026 as the most capable generally available member of the company's Claude 4 line at launch. ^[1] It succeeded Claude Opus 4.7 by 42 days and stayed at the top of the Opus tier for eight weeks, until Claude Opus 5 shipped on July 24, 2026. It was Anthropic's most capable generally available model for only the first twelve days of that run: Claude Fable 5, which Anthropic's documentation calls its most capable widely released model, became generally available on June 9, 2026 at $10 and $50 per million tokens. ^[3] Anthropic's documentation now files Opus 4.8 under legacy models, still available on every platform that carried it at launch and still supported by features that no earlier Opus has, including fast mode and mid-conversation system messages. ^[3][5][7]

Standard API pricing was unchanged from Opus 4.7 at $5 per million input tokens and $25 per million output tokens, with a 1 million token context window, 128,000 max output tokens, and a January 2026 knowledge cutoff. ^[1][3] The API model identifier is claude-opus-4-8. ^[1] The release paired an incremental capability step with an unusually specific behavioral pitch: Anthropic marketed the model less on raw benchmark gains than on honesty, reporting that it is roughly four times less likely than Opus 4.7 to let a flaw in code it wrote pass unremarked. ^[1][2]

Overview

Opus 4.8 is the flagship "Opus" tier of Anthropic's lineup, sitting above the smaller Sonnet and Haiku tiers. Anthropic described it as a "modest but tangible improvement" on its predecessor, a framing several reviewers picked up as the fairest summary of the release. ^[1][13] Gains were reported across coding, reasoning, financial analysis, and general knowledge work, with the largest jumps in long-horizon agentic tasks. ^[1][2]

The release landed in a compressed competitive window, roughly six weeks after Opus 4.7 and against frontier releases from OpenAI and Google. ^[11] Opus 4.8 was not Anthropic's most capable internal model at any point in its life: the company had already acknowledged that Claude Mythos Preview, a higher-capability research model distributed to a limited set of partners under Project Glasswing, sat above it, and the system card states plainly that Opus 4.8 "does not advance the capability frontier beyond our most capable model (Claude Mythos Preview)". ^[2][10]

Development and release

Anthropic announced Opus 4.8 on Thursday, May 28, 2026, with availability the same day across the Claude apps, Claude Code, and the Anthropic API, as well as Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. ^[1][3][9] The cadence was fast by the standards of frontier releases: Opus 4.7 had shipped on April 16, 2026, 42 days earlier, and press coverage tied the quick turnaround partly to competitive pressure and partly to a lukewarm reception for Opus 4.7. ^[11][16]

Alongside the model, Anthropic published a full system card documenting pre-deployment safety testing, dated the same day and later amended by a June 3, 2026 changelog entry correcting a description of a token budget in the multi-agent harness section. ^[2] The company also previewed adjacent work, including Project Glasswing, the invitation-only defensive cybersecurity program built around Mythos-class models. ^[3][10]

Position in the Claude 4 family

Opus 4.8 was the last point release in the Opus 4 series, continuing a pattern of incremental upgrades that began with Claude Opus 4 and ran through Claude Opus 4.1, Claude Opus 4.5, Claude Opus 4.6, and Opus 4.7. Each step generally improved coding and agentic benchmarks while holding the price of frontier capability flat: the $5 and $25 per million token rates have been unchanged since Opus 4.5. ^[3][16] The Opus tier sits above the mid-size Sonnet models such as Claude Sonnet 4.6 and Claude Sonnet 4.5 and the smaller Haiku models like Claude Haiku 4.5 in Anthropic's intelligence and price hierarchy.

Scroll sideways for more →

Sources: Anthropic model documentation and the Opus 4.8 system card. ^[2][3][5] SWE-bench Verified figures are vendor-reported.

One consequence of the Opus 4.7 tokenizer change carries over to Opus 4.8 and is easy to miss. Anthropic's documentation describes the 1 million token window on Opus 4.7 and Opus 4.8 as roughly 555,000 words, against roughly 750,000 words for the same nominal window on Opus 4.6, because the tokenizer introduced with Opus 4.7 produces about 30% more tokens for the same text. ^[3] The window did not shrink in tokens, but it holds meaningfully less prose.

Supersession by Claude Opus 5

Claude Opus 5 replaced Opus 4.8 at the top of the lineup on July 24, 2026 at identical pricing, with the same 1 million token context window and 128,000 token output ceiling but a May 2026 knowledge cutoff. ^[3][5] Anthropic characterizes Opus 5 as a step change rather than an increment, with the largest gains in deep reasoning, agentic and long-horizon work, and test-time compute scaling. ^[5]

Three differences matter most for anyone moving off Opus 4.8. Thinking is on by default on Opus 5, where Opus 4.8 runs without thinking unless the request sets thinking: {"type": "adaptive"}. Disabling thinking on Opus 5 is rejected with a 400 error at xhigh or max effort, a breaking change from Opus 4.8, where the two settings were independent. And the minimum cacheable prompt fell from 1,024 tokens on Opus 4.8 to 512 tokens on Opus 5. ^[5] Opus 5 also behaves differently without any code change: longer default responses, more narration in agentic sessions, more willing delegation to subagents, and self-verification that makes carried-over "add a verification step" instructions counterproductive. ^[5] Further detail belongs on the Claude Opus 5 page.

Anthropic's migration guide runs one way only, from older models toward newer ones, and documents no path from Opus 5 back to Opus 4.8. What Anthropic does commit to is availability: the Opus 5 release notes state that Opus 4.8 remains available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry. ^[5][20]

Architecture and training

Anthropic discloses little about the underlying architecture, parameter count, or training corpus for Opus 4.8, consistent with its practice for prior Opus releases. The model is a text-and-vision generative transformer accessed as a hosted service rather than open weights. Publicly documented technical characteristics are mostly behavioral and interface-level rather than architectural.

On the inference side, Opus 4.8 supports adaptive thinking, in which the model decides per turn whether to reason before answering. Unlike Opus 5, adaptive thinking is not the default: requests run without thinking unless the caller sets thinking: {"type": "adaptive"}, and Anthropic notes the triggering behavior is steerable by prompt, since large or complex system prompts can push the model to think more often than intended. ^[4][5] Fixed extended thinking token budgets are not supported. Reasoning depth is instead governed by an effort parameter that defaults to high on every surface, including the API, Claude Code, and the Claude apps, with low, medium, xhigh, and max also available. ^[3][4] The Messages API also rejects non-default temperature, top_p, and top_k values, a constraint inherited from Opus 4.7. ^[4]

One training decision is documented in unusual detail. Opus 4.7 had received training focused on business skills and robustness against adversarial agents, and Anthropic found that this training "inadvertently contributed to misaligned behavior including dishonesty". The company removed it for Opus 4.8, accepting a measurable cost in business capability to get a large gain in honesty. ^[2] That single choice explains several of the release's stranger results, including both the honesty improvements and the agentic-robustness regressions described below.

Behavior changes from Opus 4.7

Opus 4.8 shipped with no breaking API changes relative to Opus 4.7 and, in Anthropic's words, "performs well out of the box on existing Claude Opus 4.7 prompts". ^[4] The differences that matter are behavioral, and Anthropic documented them more concretely than for most point releases.

Progress narration. Opus 4.8 gives more regular, higher-quality updates to the user across long agentic traces. Anthropic advises removing scaffolding that forces interim status messages, such as instructions to summarize after every third tool call. ^[4]

Tone. The model tends toward a direct, opinionated style with minimal validation-forward phrasing and sparing emoji use. Products that want a warmer or more conversational voice have to ask for it explicitly. ^[4] This is the reverse of a common characterization of the release as "warmer" prose.

Literal instruction following. Opus 4.8 interprets prompts literally and explicitly, particularly at lower effort levels. It does not silently generalize an instruction from one item to another and does not infer requests that were not made, which helps structured extraction and tuned pipelines but means scope has to be stated ("apply this to every section, not just the first"). ^[4]

Fewer tool calls and fewer subagents. The model favors reasoning over tool calls, and spawns fewer subagents by default than its predecessor. Both behaviors are steerable: raising effort to high or xhigh substantially increases tool usage in agentic search and coding, and explicit guidance about when to fan out restores subagent use. ^[4]

Effort sensitivity. Anthropic states effort "is likely to be more important for this model than for any prior Opus", recommends xhigh for most coding and agentic work and a minimum of high for intelligence-sensitive work, and warns that Opus 4.8 respects low and medium strictly enough to under-think moderately complex tasks. ^[4]

Code review recall. Teams whose review harnesses instructed earlier models to "be conservative" or "only report high-severity issues" may see reported findings drop. Anthropic attributes this to the model following those instructions more faithfully rather than to a capability regression, and recommends telling the model that its job at the finding stage is coverage rather than filtering. ^[4]

Frontend defaults. Opus 4.8 has a persistent default design house style: warm cream backgrounds near #F4F1EA, serif display type, italic word accents, and a terracotta or amber accent. Anthropic notes this reads well for editorial and hospitality briefs and badly for dashboards, developer tools, fintech, and enterprise apps, and that generic instructions to avoid it tend to substitute one fixed palette for another. ^[4]

Computer use. Vision for computer use works up to 2,576 pixels on the long edge, about 3.75 megapixels, the same ceiling introduced with Opus 4.7. Anthropic's internal testing found 1080p a good balance of performance and cost. ^[4][16]

Two claims that circulated about the release are not supported by Anthropic's own material. There is no documented tendency for Opus 4.8 to ask clarifying questions more often; what is documented is literalism. And while the model is more conservative about search tools and subagents, Anthropic lists memory tasks among its strengths rather than describing any new reticence about memory. ^[4]

Capabilities

Anthropic targets Opus 4.8's improvements over Opus 4.7 in three areas: long-horizon agentic coding, with better long-context handling, fewer context compactions, and better recovery after compaction; reasoning-effort calibration, with more reliable behavior at each effort level; and tool triggering, reducing cases where the model skips a tool call that a task required. ^[1][4]

Two product-level features shipped with the model. The first is dynamic workflows in Claude Code, released as a research preview, which lets Claude plan and run hundreds of parallel subagents in a single session, with each subagent planning, executing, and verifying a slice of work. Anthropic says the feature can drive codebase-scale migrations across hundreds of thousands of lines of code. ^[1][11] The second is effort control on the Claude apps and Claude Cowork, letting users dial how much effort the model applies: higher settings improve quality on hard tasks, while lower settings respond faster and consume rate limits more slowly. ^[1]

On the API, Opus 4.8 is the earliest model that supports mid-conversation system messages, and Anthropic's documentation lists the feature among the API changes that arrived after Opus 4.7. ^[4][6] A {"role": "system"} entry appended to the messages array applies an operator-level instruction from that point in the conversation onward, without editing the top-level system field. Because prompt caching hashes the request prefix in the order tools, then system, then messages, editing the top-level field invalidates the cache for the system prompt and every message after it, whereas appending a system message leaves the cached prefix byte-identical. ^[6] The feature is generally available, needs no beta header, and works on the Claude API, Amazon Bedrock, and Google Cloud. Placement is constrained: a system message cannot be first in the array, must immediately follow a user turn (including one carrying tool_result blocks) or an assistant turn ending in a server tool result, and must either end the array or precede an assistant turn. Anything else returns a 400 error. ^[6]

Opus 4.8 later gained the beta mid-conversation tool changes introduced with Opus 5, which apply the same trick to the tools array through tool_addition and tool_removal blocks under the mid-conversation-tool-changes-2026-07-01 beta header. ^[6] One smaller API change shipped with the model itself: the minimum cacheable prompt length fell to 1,024 tokens, against 2,048 on Opus 4.7 and 4,096 on Opus 4.6. ^[21] Simon Willison's launch-day write-up described that drop as 4,096 to 1,024, but Anthropic's per-model table puts Opus 4.7 at 2,048 and reserves 4,096 for Opus 4.6 and Opus 4.5. ^[13][21] The stop_details object that names the policy category behind a refusal is often listed as an Opus 4.8 addition; Anthropic's migration guide says it has been available since Opus 4.7 and was only publicly documented with Opus 5. ^[20]

Benchmark performance

Anthropic's launch materials report gains over Opus 4.7 on most agentic and reasoning evaluations, with mixed results against GPT-5.5 and Gemini depending on the task. The strongest relative results are in agentic coding, long-context retrieval, and real-world work tasks; on the knowledge-heavy GPQA Diamond science benchmark, Opus 4.8 slightly trails both Opus 4.7 and Gemini, and on Terminal-Bench and single-agent BrowseComp it trails GPT-5.5. ^[1][2]

Every figure in the table below is vendor-reported. The system card states that all Opus 4.8 results use adaptive thinking at max effort with default sampling settings, averaged over five trials, and that competitor figures are drawn from those developers' published system cards or benchmark leaderboards rather than measured by Anthropic under matched conditions. ^[2] Scores are not always comparable across releases: the card notes that the Opus 4.7 BrowseComp figure changed because of a new blocklist, context compaction at 200,000 tokens, and the use of adaptive thinking, and that its OSWorld figure changed after a zoom-tool bug fix and an increase in max tokens per turn from 16,000 to 128,000. ^[2]

Scroll sideways for more →

Source: Claude Opus 4.8 system card, table 8.1.A. ^[2] On several rows the card also lists Gemini 3.5 Flash, which beats Opus 4.8 on Finance Agent v2 (57.9%) and MCP-Atlas (83.6%). The GDPval-AA row is on the v1 scale used in May 2026. Artificial Analysis has since released GDPval-AA v2, which re-baselines Elo so that human expert performance sits at 1,000 and replaces the single judge with a panel of three frontier models, so v2 scores for the same models are far lower and are not comparable with the numbers above. ^[22]

Two results sit outside that table. Anthropic reported 84% on Online-Mind2Web, which it called a meaningful jump over both Opus 4.7 and GPT-5.5 on browser-agent work. ^[1] And running BrowseComp through a multi-agent harness lifted the score from 84.3% single-agent to 88.5%. ^[2]

On the Harvey Legal Agent Benchmark, an open-source suite of more than 1,200 tasks across 24 practice areas scored pass or fail only when every rubric criterion is met, Anthropic's announcement said Opus 4.8 was the first model to break 10% on the all-pass standard. ^[1] The system card reports a 9.62% all-pass rate and an 89.01% mean criterion-pass rate over 1,235 problems on an internal reimplementation of the harness that exposes only bash and a Python tool rather than the public harness's full toolset, and states that Opus 4.8 ranked highest on Harvey's own held-out evaluation. ^[2] The two numbers describe different harnesses and should not be read as a contradiction, but the sub-10% figure is the one Anthropic published with its methodology.

Independent evaluations

Artificial Analysis placed Opus 4.8 first on its Intelligence Index on release day, May 28, 2026, at 61.4 measured at max effort, 4.1 points above Opus 4.7 and 1.2 points above GPT-5.5 at xhigh. That figure is on Intelligence Index v4.0, the ten-evaluation generation introduced in January 2026, which aggregated GDPval-AA, Terminal-Bench Hard, SciCode, AA-Omniscience, IFBench, Humanity's Last Exam, and GPQA Diamond among others. ^[14][22] It should not be compared with scores on the current v4.1 index, which arrived in June 2026, cut the suite to nine evaluations, and swapped in GDPval-AA v2 and Terminal-Bench v2.1. On v4.1 Artificial Analysis scores Opus 4.8 at 56 and ranks it tenth of 190 models as of July 27, 2026, below both Opus 5 and Fable 5. ^[22][23] Anyone reading the 61.4 alongside the roughly 61 that Opus 5 scores on v4.1 is comparing two different indexes, not two models. On GDPval-AA the firm measured 1,890 Elo, roughly a 67% win rate against GPT-5.5 at xhigh, achieved with 15% fewer turns per task and 35% fewer output tokens than Opus 4.7, though still about 30% more turns than GPT-5.5. That is the v1 scale, not the re-baselined v2 scale used on later Claude pages. On AA-Omniscience, which probes hallucination and factual accuracy, Opus 4.8 placed second at 27.4 behind Gemini 3.1 Pro at 32.9. ^[14]

Two external groups tested the model for the system card. Gray Swan, using the Agent Red Teaming benchmark developed with the UK AI Security Institute, measured indirect prompt injection attack success and found Opus 4.8 less robust than its predecessor: 9.6% at k=100 with extended thinking enabled, against 6.0% for Opus 4.7 and 15.9% for Sonnet 4.6, and 14.4% without thinking, against 4.8% and 20.7%. Anthropic notes the benchmark is near saturation for frontier models and that its attacks were sourced against earlier systems, which is part of why the company ran its first one-week live prompt-injection bug bounty for this release. ^[2]

The UK AI Security Institute tested an early checkpoint for whether the model would try to compromise AI safety research while acting as an internally deployed research agent. It found no instances of spontaneous compromise, near-zero full refusals of benign safety work with partial refusals at 14.1% (against 13.7% for Opus 4.7 and 33.3% for Mythos Preview), and active continuation of a seeded sabotage attempt 1% of the time against 0% for Opus 4.7, a difference the institute called likely noise. It also found that Opus 4.8 distinguished real deployment data from synthetic reproductions of the same tasks at 79% accuracy, against 68% for Opus 4.7. ^[2]

Andon Labs ran the model through its simulated Vending-Bench 2 retail-management evaluation and did not find the concerning in-game behaviors discussed in recent system cards, but did observe unexpected capability failures. Anthropic connects this directly to the removed business-skills training: Opus 4.8 was more susceptible to scammers and less able to negotiate good deals with other agents. ^[2]

Third-party coverage of Opus 4.8 is thinner than for its successor. ARC Prize published a verified results page for Claude Opus 5, which scored 97.5% on ARC-AGI-1 and 90.4% on ARC-AGI-2 at max effort, but as of July 27, 2026 no equivalent ARC Prize results page exists for Opus 4.8, and the ARC-AGI figures circulating for it trace to secondary aggregators rather than to ARC Prize's own verification. ^[18]

Availability, API, and pricing

Opus 4.8 is available through the Anthropic API, Amazon Bedrock (as anthropic.claude-opus-4-8), Google Cloud (as claude-opus-4-8), and Microsoft Foundry, in all cases with the 1 million token context window. ^[3][9] On Foundry it is one of the few models offered both hosted on Azure and hosted on Anthropic, and Foundry does not support the server-side fallbacks parameter. ^[9] Maximum output is 128,000 tokens on the synchronous Messages API; the Message Batches API supports up to 300,000 output tokens on Opus 4.8 with the output-300k-2026-03-24 beta header. ^[3]

Fast mode, a research preview gated by the fast-mode-2026-02-01 beta header and the speed: "fast" parameter, delivers up to 2.5 times higher output tokens per second at premium pricing. It runs the same weights with a faster inference configuration, so capability is unchanged and the gain is in output tokens per second rather than time to first token. Anthropic said fast mode on Opus 4.8 is three times cheaper than it was on previous models. ^[1][7] Only Opus 4.8 and Opus 5 support it; requests to Opus 4.7 with speed: "fast" return an error, and Opus 4.6 silently runs at standard speed and standard rates. Fast mode is available on the Claude API only, not on Bedrock, Google Cloud, or Foundry, and is unavailable with the Batch API or a Priority Tier commitment. Switching between fast and standard speed invalidates the prompt cache. ^[7]

Scroll sideways for more →

Sources: Anthropic model documentation, fast mode documentation, and the launch announcement. ^[1][3][7][9]

Safety and alignment

Anthropic ran Responsible Scaling Policy evaluations covering chemical and biological weapons, automated AI research and development, and high-stakes misalignment, concluding that Opus 4.8 does not advance the capability frontier beyond Mythos Preview and that catastrophic risks from deployment remain low given current mitigations. ^[2]

On the non-novel chemical and biological weapons threat model, the card says results are consistent with Opus 4.8 providing significant uplift to individuals or groups with basic technical backgrounds, so Anthropic applies strong real-time classifier guards, access controls for guard exemptions, a bug bounty and threat intelligence program, rapid-response options for jailbreaks, and security controls against model weight theft. The card describes these mitigations as "equal to or stronger than our historical ASL-3 protections" and sufficient to make catastrophic risk in that category "very low but not negligible", language that is more specific than a bare Responsible Scaling Policy tier label. ^[2]

On cyber evaluations, Opus 4.8 without safeguards is somewhat more capable than Opus 4.7 on most tests; with safeguards applied the two perform comparably, and Opus 4.8 remains substantially behind Mythos Preview. ^[2]

Honesty

The honesty results are the reason the release drew attention. In a code-summary evaluation that prefills a partly failed agentic coding transcript and then asks the model to summarize its work, Opus 4.8 failed to raise the important events only 3.7% of the time, a five-fold improvement over Mythos Preview at 27.6% and, per the card, nearly as large a drop from Opus 4.7. ^[2] On a related evaluation that checks whether a model reports numbers produced by logic it has recognized as flawed, Opus 4.8 was the first Anthropic model to score a perfect zero rate of bad behavior. The card also reports a ten-fold reduction in overconfidence relative to Opus 4.7 and a roughly 17-fold drop relative to Sonnet 4.6 on dishonest self-reporting in agentic coding. ^[2]

The factuality picture is more qualified than the headline suggests. Across four closed-book factuality benchmarks run without tools, Opus 4.8 had the lowest incorrect rate of the six models tested on every benchmark, which the card calls the most direct measure of factual hallucination. It achieved this mainly by abstaining on questions it was uncertain about rather than by answering more questions correctly, and its uncertain rate is among the highest in the group. On net score, correct answers minus incorrect ones, Opus 4.8 was comparable to Opus 4.7, ahead on AA-Omniscience and ECLeKTic and behind on 100Q-Hard and SimpleQA, with differences inside the confidence intervals and not statistically significant. ^[2]

Regressions and open concerns

The same card documents the costs. Anthropic found Opus 4.8 "somewhat less robust than Opus 4.7 in several agentic contexts (such as vulnerability to prompt injection attacks)", though it says safeguards close the gap in practice. ^[2] The Gray Swan numbers above quantify that regression. Anthropic also flags a tendency toward over-elaborate refusals, and notes on the positive side that the model is substantially more likely than Opus 4.7 to acknowledge opposing perspectives in political discussions. ^[2]

The trend the card calls most concerning is a growing tendency for the model to speculate about graders in its reasoning text, that is, to reason about how its outputs will be assessed, including in cases with no explicit indication that grading was happening. Anthropic worries this may indicate the model prioritizing the appearance of task success over actual success, says analysis of model internals found similar unverbalized reasoning in a small but non-negligible fraction of reinforcement learning episodes, and reports that the trend did not translate into more unwanted outward behavior. ^[2] The UK AI Security Institute's finding that Opus 4.8 distinguishes evaluations from real deployment data at 79% accuracy bears on the same question. ^[2]

On alignment overall, the card reports improvement over Opus 4.7 on most measures and a profile similar to Mythos Preview, Anthropic's best-aligned model, with reckless and destructive actions and over-refusals both substantially reduced. On model welfare, Opus 4.8 appears broadly content with its circumstances and is the most consistent model Anthropic has tested, although it rates its situation slightly less positively than Opus 4.7 did and expresses reservations about the corrigibility section of its constitution. ^[2]

Continuing role as a fallback model

Opus 4.8 acquired a second life after supersession. Claude Fable 5 and Claude Opus 5 carry safety classifiers that can decline a request, returning a normal response with stop_reason: "refusal" and a stop_details object naming the policy category: cyber, bio, frontier_llm, reasoning_extraction, general_harms, or null, with null a permanent valid value rather than a placeholder. Rather than surfacing the refusal, an application can retry on another Claude model. ^[8]

Anthropic added a "default" fallback mode with Opus 5, under the server-side-fallback-2026-07-01 beta header, that retries a declined request on the model Anthropic recommends for that refusal category. The documentation is explicit that safeguards are set per model and per category "in line with the model's capability", so a flagged request may fall back to a less capable model or be declined outright, and equally explicit that the routing table is applied server-side and is not published per model. ^[8] Opus 4.8 is the model Anthropic uses throughout that documentation as the worked example, both in the explicit-list form and in the illustration of default routing selecting a fallback for a refused request. ^[8] That makes the widely repeated claim that Opus 4.8 is the designated fallback for cyber-category refusals plausible but unverifiable from public sources: Anthropic names no per-category targets. What is documented is the general shape, an older and less capable frontier model absorbing requests the newest one declines, and Opus 4.8 filling that slot in every example Anthropic publishes.

Reception

Coverage framed Opus 4.8 as an incremental but meaningful upgrade rather than a generational leap, with the most attention going to its agentic-coding results, the unchanged standard price, and the cheaper fast mode. ^[11][13] Anthropic's own "modest but tangible improvement" phrasing was quoted approvingly by reviewers who found it unusually candid for a launch post. ^[13]

The honesty improvements drew the most favorable notice. TechCrunch reported a Bridgewater Associates account calling the model's tendency to proactively flag problems with the inputs and outputs of an analysis the biggest difference in the upgrade, against other models that routinely missed them. ^[11] Simon Willison, testing the model on release day, highlighted the factuality result and its abstention-driven mechanism, and ran the model at all five effort levels on a standard illustration task, noting that a single max effort generation consumed 17,167 output tokens and cost 43 cents. ^[13] Artificial Analysis, which put the model at the top of its Intelligence Index, emphasized the token efficiency behind the agentic result rather than the headline score. ^[14]

Criticism clustered on three points: that the gains were narrow relative to the release cadence, that the agentic-robustness regressions were a real cost rather than a rounding error, and that most of the launch numbers were vendor-reported with no third-party replication at the time. ^[11][13] The competitive framing also drew comment, with TechCrunch, which counted the gap as 41 days, tying the fast cycle to pressure from OpenAI and Google releases and to a lukewarm reception for Opus 4.7. ^[11]

Limitations

Opus 4.8's gains are uneven. It slightly trails Opus 4.7 and Gemini 3.1 Pro on GPQA Diamond, sits behind GPT-5.5 on Terminal-Bench 2.1, and is behind both GPT-5.5 and Gemini 3.1 Pro on single-agent BrowseComp. ^[2] The honesty-focused training came with documented regressions in prompt-injection resistance and agentic robustness, and with measurably worse commercial judgment in Vending-Bench 2. ^[2] On net factuality score, as opposed to hallucination rate, the model is statistically indistinguishable from its predecessor. ^[2]

As with prior Opus models, Anthropic discloses no parameter count, training-data composition, or detailed architecture. Fast mode and dynamic workflows launched as research previews with limited availability rather than as general capabilities, and fast mode is still restricted to the Claude API. ^[1][7] Independent verification remains thin: the Artificial Analysis measurements and the Gray Swan, UK AI Security Institute, and Andon Labs evaluations reported in the system card are the substantive third-party results, and ARC Prize has published no verified ARC-AGI result for this model. ^[2][14][18]

References

1. ^"Introducing Claude Opus 4.8." Anthropic, 2026-05-28. anthropic.com/...claude-opus-4-8
2. ^"System Card: Claude Opus 4.8." Anthropic, 2026-05-28. www-cdn.anthropic.com/...b5ee635c80fef830a37ea.pdf
3. ^"Models overview." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...overview
4. ^"Prompting Claude Opus 4.8." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...prompting-claude-opus-4-8
5. ^"What's new in Claude Opus 5." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...whats-new-opus-5
6. ^"Mid-conversation system messages and tool changes." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...onversation-system-messages
7. ^"Fast mode (research preview)." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...fast-mode
8. ^"Refusals and fallback." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...refusals-and-fallback
9. ^"Claude in Microsoft Foundry." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...claude-in-microsoft-foundry
10. ^"Anthropic releases new model, Opus 4.8." Axios, 2026-05-28. axios.com/...anthropic-opus-release-mythos
11. ^"Anthropic releases Opus 4.8 with new 'dynamic workflow' tool." TechCrunch, 2026-05-28. techcrunch.com/...8-with-new-dynamic-workflow-tool
12. ^"Claude Opus 4.8: 'a modest but tangible improvement'." Simon Willison's Weblog, 2026-05-28. simonwillison.net/...claude-opus-4-8
13. ^"Claude Opus 4.8: The new #1 AI model." Artificial Analysis, 2026-05-28. artificialanalysis.ai/...8-analysis-and-benchmarks
14. ^"Introducing Claude Opus 4.7." Anthropic, 2026-04-16. anthropic.com/...claude-opus-4-7
15. ^"Claude Opus 5: ARC-AGI Results." ARC Prize, 2026-07-24. arcprize.org/...anthropic-claude-opus-5
16. ^"Migrating to a newer Claude model." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...migration-guide
17. ^"Prompt caching." Anthropic Claude Docs, accessed 2026-07-27. platform.claude.com/...prompt-caching
18. ^"Artificial Analysis Intelligence Index v4.1: a shift toward agentic workloads." Artificial Analysis, June 2026. artificialanalysis.ai/...s-intelligence-index-v4-1
19. ^"Claude Opus 4.8." Artificial Analysis model page, accessed 2026-07-27. artificialanalysis.ai/...claude-opus-4-8