Model Drop

Claude Opus 5

July 24, 2026

Happy 20th edition of Model Drop! Today: Claude Opus 5, which Anthropic ships just two months after Opus 4.8 and six weeks after Fable 5. That’s four frontier models from one lab in under two months.

Model: Claude Opus 5 (claude-opus-5 on the Claude API and Google Cloud, anthropic.claude-opus-5 on Bedrock)

Model type: Text + image input, text output. Vision and multilingual. No native image, audio, or video output.

Ship date: July 24, 2026

Maker: Anthropic (San Francisco)

Pricing: $5 / $25 per million input / output tokens, identical to Opus 4.8 and half of Fable 5’s $10 / $50. Standard prompt caching and Batch API discounts apply, and the minimum cacheable prompt drops to 512 tokens from 1,024. Fast mode is a research preview at $10 / $50 for roughly 2.5x the default speed (Claude API only)

Available on: claude.ai and the Claude apps (the new default on Claude Max, the strongest model available on Claude Pro), Claude Code, Claude Cowork, the Claude Developer Platform, Claude in Amazon Bedrock, Claude on Google Cloud, and Microsoft Foundry

Headline benchmarks: Anthropic published ratios rather than raw scores almost everywhere. On Frontier-Bench v0.1 it says Opus 5 “surpasses all other models, and more than doubles Opus 4.8’s performance at a lower cost per task.” On CursorBench 3.2 at max effort it lands “within 0.5% of Fable 5’s peak score, but at half the cost per task.” On ARC-AGI 3 its score is “three times as high as the next-best model” (a commenter on the launch thread pegs that at 30% against GPT-5.6’s 7.8%). On OSWorld 2.0 it beats Fable 5’s best result “at just over a third of the cost.” Zapier’s AutomationBench pass rate is “around 1.5x the next-best model for the same cost per task.” In the life sciences it’s 10.2 points over Opus 4.8 on organic chemistry and 7.7 points on protein tasks.

Other info: 1M-token context window (both the default and the maximum, with no smaller variant), 128K max output, 300K through the Batch API extended-output beta. Reliable knowledge cutoff and training cutoff both May 2026, the freshest of any Claude model. Full low / medium / high / xhigh / max effort ladder with no beta header required. Anthropic calls it the most aligned model it has shipped, scoring 2.3 on overall misaligned behavior and adhering to Claude’s Constitution better than Opus 4.8, Sonnet 5, or Fable 5. Safeguards match Fable 5’s with one loosening: source-code vulnerability discovery is permitted at every access level, while binary scanning and exploit generation stay blocked. Cyber classifiers are expected to fire about 85% less often than they do for Fable 5. Unlike Fable 5 and Mythos 5, Opus 5 carries no 30-day data retention requirement for general access. System card published at launch.

More details: Introducing Claude Opus 5 and the Claude Opus 5 System Card.

Anthropic released Claude Opus 5 today, making it the default on Claude Max and the ceiling on Claude Pro the same day, with the API live across Anthropic’s own platform, AWS, Google Cloud, and Microsoft Foundry. The pitch is one sentence long: “a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.” Price is the whole story, as is becoming the case with many frontier launches. Opus 5 costs exactly what Opus 4.8 cost, $5 in and $25 out, while Anthropic claims it more than doubles 4.8 on its own frontier coding eval and lands within half a point of Fable 5 on CursorBench. Fortune reports that business customers had been complaining about Fable 5’s burn rate, blowing through token budgets and running up bills, so voila: Opus 5.

The supporting evidence is broad and oddly shy about absolute numbers. Anthropic leads with ratios: 3x the next-best model on ARC-AGI 3, 1.5x the pass rate on Zapier’s AutomationBench at the same cost per task, better than Fable 5 on OSWorld 2.0 at a third of the cost, more than double Opus 4.8 on Frontier-Bench. Launch partners backed the framing, with Cognition’s Scott Wu saying Opus 5 “approaches Fable-level performance at half the cost” and Cursor’s Sualeh Asif calling it “near Fable 5 intelligence at Opus speed and cost.” It got 2.3 on Anthropic’s automated misaligned-behavior audit, the lowest of any recent Claude, plus a meaningful loosening of the cyber classifiers that had been strangling legitimate security work on Fable 5, and no 30-day retention requirement. TechCrunch noted that Opus 5 is a smaller model than Fable 5 yet beats it on a number of benchmarks, which raises the obvious question of what Fable 5 is now for. Anthropic’s own answer is that Fable still wins on multi-day autonomous projects and stays ahead on biology research and offensive cyber. And the docs quietly warn that Opus 5’s default responses run longer than prior Opus models (which means more money, not less).

What’s new

Opus 5 is the first Claude release in a while where the API changes matter as much as the benchmark chart.

A max effort tier, no beta header. The full effort ladder runs low through max, with max as the explicit top of the range for the deepest reasoning. Anthropic specifically calls out test-time compute scaling as one of the biggest gains, meaning extra effort actually converts into better answers rather than more tokens.
Thinking on by default, and a breaking change under it. On Opus 4.8 you opted into thinking. On Opus 5 it’s on unless you turn it off, and you can only turn it off at effort high or below. Request xhigh or max with thinking: {"type": "disabled"} and you get a 400. Anthropic also warns that with thinking off the model will occasionally write a tool call into its text output or leak internal XML tags (which is a polite way of saying don’t do that).
Mid-conversation tool changes. You can now add or remove tools between turns without blowing the prompt cache, instead of shipping one frozen tool list for the life of a session. For anyone building agents where the toolset changes as the task moves through phases, this removes a real and annoying tax.
Automatic fallbacks for refusals. The fallbacks parameter gets a "default" mode that routes safety-flagged requests to Anthropic’s recommended fallback model by refusal category, so a tripped classifier returns a usable answer instead of an error. Combined with cyber classifiers firing 85% less often than Fable 5’s, this is Anthropic responding directly to the complaint that its safest models were unusable.
It verifies its own work now. The docs tell you to delete verification instructions carried over from older models (”include a final verification step,” “use a subagent to verify”) because they cause over-verification on Opus 5. It also narrates progress more and delegates to subagents more readily. Rare to see a model launch ship with a list of prompts you should remove.

How and where to use it

Where it runs, what it’s for, and when to reach for something else.