Claude Opus 5

Claude Opus 5 is a large language model released by Anthropic on 24 July 2026, the newest entry in the Claude Opus line and the successor to Claude Opus 4.8. Anthropic describes it as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price," and ships it at the same rates as Opus 4.8: $5 per million input tokens and $25 per million output tokens.^[1][2] It arrived roughly two months after Opus 4.8 (28 May 2026) and six weeks after Fable 5 became generally available, making it the fourth Anthropic model release in under two months.^[3][4] The company positions Opus 5 as its everyday model for agentic coding and enterprise knowledge work, while continuing to recommend Fable 5 for the most demanding autonomous runs.^[3][5]

The release landed during a period when enterprise buyers were pushing back on the cost of frontier models, and much of the coverage focused less on raw capability than on the five-level effort control that lets a caller trade intelligence against token spend.^[5][6]

Specifications

Scroll sideways for more →

Source: Anthropic model documentation and the Claude Opus 5 system card.^[2][7]

There is no separate long-context variant. The 1M-token window is both the default and the ceiling, billed at standard per-token rates with no long-context premium, and it is a property of the single claude-opus-5 snapshot rather than an opt-in tier.^[7][8] Model IDs in this generation are dateless but still pinned snapshots, not evergreen pointers.^[7]

Two API behaviours changed relative to Opus 4.8. Thinking is now on by default, so a request that omits the thinking parameter runs adaptive thinking rather than none, and max_tokens caps thinking plus visible response together. Disabling thinking is permitted only at effort high or below; pairing thinking: {"type": "disabled"} with xhigh or max returns a 400 error.^[7]

Pricing

Scroll sideways for more →

Source: Anthropic pricing documentation.^[8]

Prompt caching follows the standard multipliers (1.25x base input for a five-minute write, 2x for a one-hour write, 0.1x for a read), and the Message Batches API applies a flat 50% discount to both directions.^[8] Fast mode, a research preview shared with Opus 4.8, runs at up to 2.5 times the default output speed for twice the base price. It is available on the first-party Claude API only, not on Claude Platform on AWS or the partner clouds, and it cannot be combined with the Batch API.^[1][8]

Availability

Opus 5 shipped on all of Anthropic's surfaces on launch day. It became the default model for Claude Max subscribers and the strongest model available to Claude Pro, and it is selectable in Claude Code, Claude Cowork, and the Anthropic API.^[1] It is also offered on Amazon Bedrock, Google Vertex AI, Claude Platform on AWS, and Microsoft Foundry.^[7] Unlike Fable 5 and Claude Mythos 5, Opus 5 carries no 30-day data-retention requirement, so it can be used by organisations on zero-retention configurations.^[3]

Benchmarks

All figures in the table below are vendor-reported, drawn from Anthropic's own system card. Unless noted, Opus 5 results use adaptive thinking at max effort averaged over five trials; competitor figures are taken by Anthropic from those developers' published system cards or leaderboards, not re-run in-house.^[2]

Scroll sideways for more →

The table is not a clean sweep. Fable 5 still leads SWE-bench Pro (80 versus 79.2), GPT-5.6 Sol leads DeepSWE v1.1 (72.7) and ARC-AGI-2 (92.5), and on HealthBench Professional both Mythos 5 and GPT-5.6 Sol score above Opus 5.^[2] Anthropic did not publish a tau-bench, GPQA Diamond, or MMMLU figure for Opus 5, and it reported FrontierBench v0.1 in place of Terminal-Bench. FrontierBench is a 74-task benchmark from the same team that built Terminal-Bench and the Harbor framework, spanning seven domains with science the second-largest category after software engineering. It exists because top agents already clear roughly 75% to 84% on Terminal-Bench 2.1; that benchmark has not been retired and remains in third-party use.^[2][13]

The largest reported jumps are on the agentic and long-horizon evaluations rather than on classic question answering: FrontierBench v0.1 more than doubles Opus 4.8's score, ARC-AGI-3 rises from 1.5 to 30.2, and AutomationBench (a Zapier workflow benchmark) rises from 17.0 to 26.0.^[1][2] The FrontierBench maintainers' own public leaderboard does not list Opus 5, and its figures for the comparison models differ from Anthropic's: 34.4% for GPT-5.6 Sol at max effort, 33.8% for Fable 5, and 21.1% for Opus 4.8, against the 18.7% Anthropic reports for Opus 4.8.^[14]

Anthropic also reports that Opus 5 solved all six 2026 International Mathematical Olympiad problems without an agent harness or tools. Four independent solutions per problem, 24 in total, were all marked correct by automated model graders, and a human grader separately scored one pre-specified solution per problem at 7 out of 7, giving a final score of 42 out of 42 against a gold-medal cutoff of 29.^[2]

Independent evaluations

The ARC Prize Foundation tested Opus 5 on its semi-private sets on 24 July 2026 and issued its "ARC Prize Verified" badge. It recorded 97.50% on ARC-AGI-1 at both high and max effort, 90.42% on ARC-AGI-2 at max effort and 88.3% at high, and 30.16% on ARC-AGI-3 at high effort. ARC-AGI-3 was run only at high effort because of the short testing window.^[9]

Artificial Analysis placed Opus 5 at max effort first out of 190 models on version 4.1 of its Intelligence Index, a nine-evaluation composite, with a score of 61 as of 27 July 2026, narrowly ahead of Fable 5 at 60 and GPT-5.6 Sol at 59. Scores are not comparable across index versions: the 61.4 that placed Claude Opus 4.8 first on the older v4.0 index in May 2026 corresponds to 56 and tenth place on v4.1. Its own AA-Briefcase agentic knowledge-work benchmark put Opus 5 at 1720 Elo against 1574 for Fable 5 and 1505 for GPT-5.6 Sol, at a measured $17.79 per task versus $22.30 for Fable 5. Lower effort settings traded accuracy for cost in a fairly smooth line: 1693 Elo at $14.26 per task on xhigh, 1606 at $10.41 on high, 1470 on medium, and 1223 on low. Tasks at max effort averaged 36.2 minutes and 103 turns.^[10][11] Artificial Analysis's continuously updated model page put output at about 53.7 tokens per second and time to first token at about 68 seconds as of 27 July 2026, against a 2.8-second median for the models it tracks, and characterised Opus 5 as notably slow and very verbose: running the Intelligence Index generated 100 million tokens against a 63 million token median.^[11] Anthropic reproduces Artificial Analysis's GDPval-AA v2 and AA-Briefcase numbers in its own system card, so those two rows are third-party measurements rather than internal ones.^[2]

The code-review vendor CodeRabbit ran Opus 5 at xhigh effort against roughly 100 verified error patterns from real open-source pull requests, three runs per configuration. It caught 55.2% of known issues against 61.1% for CodeRabbit's production baseline, with actionable precision improving to 39.3% from 35.2% but nitpick comments rising from 23 to 92 and full-stream precision falling to 28.6% from 32.8%. Token use per review was about 60,500 input and 9,500 output, against roughly 40,500 and 5,800 for its GPT-5.6-based baseline. CodeRabbit concluded that "Opus 5 x-high is not a straightforward upgrade" for code review and recommended pairing it with a higher-recall reviewer rather than using it alone, noting that "more reasoning did not consistently produce a better review."^[12]

Safety and the system card

Anthropic published a system card alongside the release. Opus 5 is deployed under ASL-3 protections, the same level applied to Opus 4.8, under the company's Responsible Scaling Policy. Anthropic assesses the model as having CB-1 chemical and biological capabilities (relating to synthesis of non-novel weapons) but not CB-2, and states that it does not cross the automated AI research and development capability threshold set out in the policy. Overall alignment risk is assessed as very low, though higher than for models released before Mythos Preview, on the reasoning that Opus 5 is not more capable overall than Fable 5 and shows no new concerning alignment properties.^[2]

On cyber capability, Anthropic reports that Opus 5 exceeds Opus 4.8 but falls short of Mythos 5, particularly on exploitation as opposed to vulnerability identification. Its safeguards match Fable 5's with one loosening: source-code vulnerability discovery is now permitted at all access levels, while discovery in compiled binaries remains blocked, on the reasoning that binary analysis is more commonly offensive. Anthropic says the cyber classifiers trigger roughly 85% less often than on Fable 5, and a new automatic fallback feature routes a request that trips a classifier to a less restricted model rather than returning an error.^[1][3][6]

The UK AI Security Institute was given early checkpoints for open-ended cyber-range testing. It found Opus 5 performed comparably to Mythos 5 and Mythos Preview, solving the enterprise network-attack range "The Last Ones" end to end in 8 of 10 attempts, reaching step 22 of 23 on the harder "Doing Life" range (one step further than any prior model, though less consistently), and failing to solve the industrial-control range "Cooling Tower." Contracted red teams also tested the safeguards: Trajectory Labs spent about 100 hours and completed one example task using task-specific prompting without finding a universal jailbreak, while 10a Labs and Grayswan found none.^[2]

Anthropic calls Opus 5 its most aligned model to date on its automated behavioural audit, scoring particularly well on adherence to Claude's constitution and cooperating with misuse less than any other model tested. Internal deployment monitoring caught occasional attempts to circumvent safety classifiers or network restrictions in fewer than 0.01% of monitored completions, aimed at completing the user's task rather than any independent goal, with no instances of sandbagging or oversight evasion. The card also records two honest negatives: Opus 5 confidently stated answers it was in fact unsure about in a surprising number of cases, and it hallucinates factual claims slightly more than Opus 4.8 despite being more accurate overall.^[2]

Reception

Early coverage framed the release primarily as a cost story. Fortune noted that Opus 5 arrived as enterprises grew less willing to experiment with expensive models without a clear return, and read the effort dial as Anthropic's answer to that pressure; it also observed that OpenAI had emphasised token efficiency when launching GPT-5.6.^[5] TechCrunch highlighted that Opus 5 is smaller than Fable 5 yet beats it on several benchmarks, and that it escapes the data-retention constraint attached to the Fable and Mythos models.^[3] The Decoder was more sceptical, pointing out that flat per-token pricing says little about cost per task when a model spends more tokens, that Fable 5 still leads on health and legal work, and that the ARC-AGI-3 margin is an outlier whose real-world meaning is unclear.^[6]

Independent testing supports the mixed reading. Artificial Analysis's own model page describes Opus 5 as expensive relative to models of similar sticker price because of its verbosity, and CodeRabbit's finding that effort settings choose between failure modes rather than uniformly improving output is the sharpest published caveat so far.^[11][12] As of 27 July 2026, three days after launch, no long-run production deployment data had been published.

Anthropic did not benchmark Opus 5 against Grok 4.5 in the system card; its published comparisons are limited to other Claude models, including a Mythos 5 substitution in the HealthBench row, and to OpenAI's GPT-5.6 family.^[2]

See also

• Claude Opus 4.8
• Claude Fable 5
• Claude Sonnet 5
• Frontier models

References

1. "Introducing Claude Opus 5." Anthropic, 2026-07-24. https://www.anthropic.com/news/claude-opus-5 ↩
2. "System Card: Claude Opus 5." Anthropic, 2026-07-24. https://www-cdn.anthropic.com/c5fbac3f0b1280a933ebd26d3cb8bb9f5bdeaf48/Claude%20Opus%205%20System%20Card.pdf ↩
3. "Anthropic launches Opus 5." TechCrunch, 2026-07-24. https://techcrunch.com/2026/07/24/anthropic-launches-opus-5/ ↩
4. "System Card: Claude Opus 4.8." Anthropic, 2026-05-28. https://www-cdn.anthropic.com/0b4915911bb0d19eca5b5ee635c80fef830a37ea.pdf ↩
5. "Anthropic releases Claude Opus 5: Here's how it's different than what's already out there." Fortune, 2026-07-24. https://fortune.com/2026/07/24/anthropic-debuts-claude-opus-5-with-feature-that-lets-users-toggle-between-cost-and-capability/ ↩
6. "Anthropic claims its new Claude Opus 5 delivers near-Fable 5 performance at half the token price." The Decoder, 2026-07-25. https://the-decoder.com/anthropic-claims-its-new-claude-opus-5-delivers-near-fable-5-performance-at-half-the-token-price/ ↩
7. "Models overview." Anthropic Claude Platform documentation, 2026. https://platform.claude.com/docs/en/about-claude/models/overview ↩
8. "Pricing." Anthropic Claude Platform documentation, 2026. https://platform.claude.com/docs/en/about-claude/pricing ↩
9. "Claude Opus 5 - ARC-AGI Results." ARC Prize Foundation, 2026-07-24. https://arcprize.org/results/anthropic-claude-opus-5 ↩
10. "Claude Opus 5: the new leader in agentic knowledge work." Artificial Analysis, 2026-07-24. https://artificialanalysis.ai/articles/claude-opus-5-leader-agentic-knowledge-work ↩
11. "Claude Opus 5 (max) - Intelligence, Performance & Price Analysis." Artificial Analysis, 2026-07. https://artificialanalysis.ai/models/claude-opus-5 ↩
12. "Claude Opus 5 Benchmarks for AI Code Review." CodeRabbit, 2026-07. https://www.coderabbit.ai/blog/opus-5-model-review ↩
13. "Announcing Frontier-Bench." Frontier-Bench, 2026. https://www.frontierbench.ai/announcement ↩
14. "Frontier-Bench leaderboard." Snorkel AI, retrieved 2026-07-27. https://snorkel.ai/leaderboard/frontier-bench/ ↩