Image: Anthropic / anthropic.com Anthropic Ships Claude Opus 5: 96% on SWE-bench, a Gold Medal on IMO 2026, and a 1M Context Window
Claude Opus 5 arrived July 24 with the same pricing as Opus 4.8, a 1 million-token context window, and benchmark numbers that push past Claude Fable 5 on agentic and computer-use tasks. It scores 96% on SWE-bench Verified, solves every problem on IMO 2026, and includes a new effort dial that lets you trade cost for capability per request.
Anthropic released Claude Opus 5 on July 24. The model ID is claude-opus-5. Pricing stays the same as Opus 4.8: $5 per million input tokens, $25 per million output tokens. Fast mode is available at 2x the base price and runs roughly 2.5x faster.
The context window goes to 1 million tokens, matching what Anthropic already offered on Fable 5. Max output is 128k tokens on the Messages API, or up to 300k via the Batches API with a beta header.
The benchmark numbers
On SWE-bench Verified, Opus 5 scores 96%. That’s a significant jump and puts it at the top of published scores. On SWE-bench Pro, it reaches 79.2%, up from 69.2% for Opus 4.8.
The agentic results are striking. On OSWorld 2.0, a benchmark for computer-use tasks, Opus 5 scores 70.57%, up from 55.7% for Opus 4.8, and outperforms Claude Fable 5 at roughly one-third the cost. On Zapier AutomationBench, it reaches 26%, up from 17% for Opus 4.8.
ARC-AGI-3, a test of general reasoning and novel problem-solving, is where the numbers get unusual. Opus 5 scores 30.16% at high effort, which Anthropic says is four times the previous best result.
On FrontierBench v0.1, a 74-task agentic benchmark, Opus 5 scores 43.3% at max effort. Opus 4.8 scored 18.7%.
Outside of code and automation, Opus 5 solved all 42 problems on the 2026 International Mathematical Olympiad, achieving a gold-medal-level score. On Humanity’s Last Exam, it scores 56.3% without tools and 64.7% with tools.
The effort dial
Opus 5 introduces a five-level effort parameter (low, medium, high, xhigh, max) that controls how much compute the model applies to a request. At lower effort levels, Anthropic says it preserves much of the performance while using fewer tokens. You can set it per request, which lets you tune cost against capability on a task-by-task basis rather than picking one model and living with its economics across everything.
One breaking change: setting thinking: {"type": "disabled"} alongside xhigh or max effort now returns a 400 error. Thinking is on by default at higher effort levels, and disabling it with those effort settings is treated as an invalid combination. Developers need to either remove the disable setting or lower the effort level.
What’s new beyond benchmarks
Two API features arrived in beta alongside the model.
Mid-conversation tool changes let you add, remove, or swap tools between turns without abandoning the session. Previously, changing the available tools required starting a new conversation. This matters for long-running agent workflows where the set of relevant tools shifts as the task progresses.
Automatic API fallbacks let you configure the API to fall back to a secondary model if the primary is unavailable, at capacity, or returns an error above a threshold. It’s opt-in and aimed at production pipelines where Opus 5 is the preferred model but reliability matters more than always using the exact model.
Safety
Prompt injection resistance in browser environments improved significantly. In Anthropic’s testing, the attack success rate dropped from 31.5% to 3.7% in browser contexts. Misaligned behavior across safety evaluations scores 2.3, the lowest Anthropic has reported for an Opus model.
Opus 5 is now the new default model on Claude Max, and the strongest model available on Claude Pro.
The full announcement is at anthropic.com/news/claude-opus-5.