Claude Opus 5 is Anthropic's new flagship large language model, released on July 24, 2026. It costs $5 per million input tokens and $25 per million output tokens — the same price as Opus 4.8 — while roughly doubling Opus 4.8's score on frontier software-engineering benchmarks and approaching Anthropic's most capable model, Fable 5, on coding and knowledge work. If you already build on Claude, Opus 5 is a drop-in upgrade at no extra cost per token.
That pricing detail is the headline. In a year where every frontier launch has pushed prices up or added a "reasoning premium," Anthropic shipped a materially smarter model at a flat rate. According to Anthropic's release notes, Opus 5 is now the default model on Claude Max and ships with a 1M-token context window, thinking on by default, and a knowledge cutoff of May 2026 — the most recent of any Claude model to date.
This guide breaks down what actually changed: the benchmark numbers, the context and pricing details, how Opus 5 compares to Opus 4.8 and GPT-5.6, and whether it's worth switching today.
Key Takeaways
- Claude Opus 5 launched July 24, 2026 at $5/$25 per million input/output tokens — identical to Opus 4.8.
- On Frontier-Bench v0.1 (software engineering), Opus 5 takes the top spot and doubles Opus 4.8's score at a lower cost per task.
- It offers a 1M-token context window (default and max), 128K max output, and thinking on by default.
- Knowledge cutoff is May 2026, and Anthropic calls it their most aligned model with the lowest rates of deceptive behavior.
- A Fast mode runs roughly 2.5x the default speed at twice the base price for latency-sensitive work.
What is Claude Opus 5?
Claude Opus 5 is the newest model in Anthropic's Opus tier, positioned for the most demanding coding, agentic, and knowledge-work tasks. It replaces Opus 4.8 as the default on Claude Max and is available across the API, Claude apps, and Anthropic's coding tool. Anthropic reports it approaches Fable 5 — its most capable widely released model — on coding and knowledge benchmarks, while costing a fraction of Fable 5's price.
The practical framing: Opus 5 narrows the gap between the "workhorse" Opus tier and the premium Fable tier. For most teams, that means near-top-of-market intelligence without paying top-of-market token rates. If you want the broader landscape first, our guide to the best AI coding assistants in 2026 puts Opus 5's competitors in context.
How good are Claude Opus 5's benchmarks?
Claude Opus 5's benchmarks are its strongest selling point: it takes the top spot on Frontier-Bench v0.1 for software engineering and doubles Opus 4.8's score at a lower cost per task. It also scores three times as high as the next-best model on ARC-AGI 3, a test of abstract reasoning, and surpasses Fable 5's best OSWorld 2.0 result at just over a third of the cost.
Here is how the headline numbers stack up, based on figures reported at launch:
| Benchmark | What it measures | Opus 5 result |
|---|---|---|
| Frontier-Bench v0.1 | Real-world software engineering | Top spot; ~2x Opus 4.8, lower cost/task |
| ARC-AGI 3 | Abstract reasoning / generalization | ~3x the next-best model |
| OSWorld 2.0 | Computer-use / agentic tasks | Beats Fable 5's best at ~1/3 the cost |
Two things stand out. First, the software-engineering gain is not marginal — doubling a frontier SWE score between minor version bumps is unusual, and it lands at a lower cost per completed task because Opus 5 resolves problems in fewer tokens. Second, the ARC-AGI 3 result matters because that benchmark resists memorization; a 3x lead suggests genuine reasoning improvement, not benchmark contamination.
A caveat worth stating plainly: these are launch figures, and vendor-run benchmarks always flatter the vendor. Independent evaluations from sites like BenchLM and community testing will refine the picture over the coming weeks. But the direction is clear and consistent across three independent test families.
Claude Opus 5 pricing and context window
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, with a 1M-token context window that is both the default and the maximum. Max output is 128K tokens, and thinking (extended reasoning) is on by default, so you don't need to configure a thinking budget.
| Spec | Claude Opus 5 |
|---|---|
| Input price | $5 / million tokens |
| Output price | $25 / million tokens |
| Context window | 1M tokens (default = max) |
| Max output | 128K tokens |
| Thinking | On by default |
| Knowledge cutoff | May 2026 |
| Fast mode | ~2.5x speed at 2x base price |
The flat pricing is the strategic move. Anthropic held the line at Opus 4.8 rates while shipping a meaningfully smarter model, which effectively lowers the cost-per-solved-task across the board. For agentic workloads that burn output tokens, the fewer-tokens-per-task efficiency compounds the savings. For latency-sensitive products, Fast mode trades price for speed at roughly 2.5x throughput.
Is Claude Opus 5 better than Opus 4.8 and GPT-5.6?
Yes — Claude Opus 5 beats Opus 4.8 decisively on coding, reasoning, and agentic benchmarks at the same price, making the upgrade a clear win for existing Anthropic users. Against OpenAI's GPT-5.6 family, Opus 5 is competitive on coding and leads on abstract reasoning, though head-to-head verdicts depend on your specific workload.
- vs Opus 4.8: Same price, roughly double the software-engineering score, more current knowledge (May 2026 cutoff), and better alignment. There is no reason to stay on 4.8 for new work unless you've pinned a version for reproducibility.
- vs Sonnet 5: Opus 5 is the heavier, more capable tier. If your workload runs comfortably on the cheaper Claude Sonnet 5, stay there; reach for Opus 5 when tasks are genuinely hard or long-horizon.
- vs GPT-5.6: OpenAI's flagship, covered in our GPT-5.6 Sol, Terra, and Luna explainer, remains strong on agentic coding. Opus 5's edge is abstract reasoning (ARC-AGI 3) and its flat, predictable pricing.
In practice, the biggest reason to choose Opus 5 is workflow: if your team already uses Claude in an editor or agent, Opus 5 is a config change, not a migration. Anthropic's own coding tool — which we cover in why Claude Code became the #1 AI coding tool — now defaults to the new model.
Alignment and safety improvements
Anthropic describes Claude Opus 5 as its most aligned model to date, with the lowest measured rates of deceptive behavior of any Claude release. For enterprise buyers, that is not a soft metric — deceptive or reward-hacking behavior in an autonomous agent is a direct operational risk, and a model that scores lower on those tests is one you can trust with longer, less-supervised runs.
This matters most for agentic deployments. A model that runs for 20 minutes unsupervised, edits files, and reports its own progress needs to report honestly. Anthropic's emphasis on grounded progress claims and reduced deception is aimed squarely at that use case. When you're wiring Opus 5 into a build pipeline or research agent, the alignment gains translate into fewer fabricated "done" reports and less silent failure.
How to start using Claude Opus 5
Getting started with Claude Opus 5 takes one change: point your API calls or app settings at the new model. On Claude Max it is already the default, and in the API you select it the same way you'd select any current Claude model.
- Update your model string in API calls to the Opus 5 identifier, or select it in the model picker.
- Leave thinking on — it's the default and tuned for hard tasks; only disable it for fast, simple calls.
- Set generous output limits for agentic work so long responses don't truncate mid-task.
- Consider Fast mode if your product is latency-sensitive and can absorb the 2x price for ~2.5x speed.
- Re-baseline your costs — Opus 5 often resolves tasks in fewer tokens, so measure real spend rather than assuming it matches 4.8.
Because pricing and the request surface match Opus 4.8, most existing integrations work unchanged. Test one representative workload, compare quality and token spend, then roll out.
Frequently asked questions
When was Claude Opus 5 released?
Claude Opus 5 was released on July 24, 2026. It immediately became the default model on Claude Max and is available through the Anthropic API and Claude apps.
How much does Claude Opus 5 cost?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8. A Fast mode runs at roughly 2.5x the default speed for twice the base price.
What is Claude Opus 5's context window?
Claude Opus 5 has a 1M-token context window, which is both the default and the maximum. It supports up to 128K output tokens per response, with extended thinking enabled by default.
Is Claude Opus 5 better than GPT-5.6?
Claude Opus 5 leads on abstract reasoning (ARC-AGI 3) and is competitive with GPT-5.6 on coding, while offering flat, predictable pricing. The best choice depends on your specific workload and existing tooling.
Should I upgrade from Opus 4.8 to Opus 5?
Yes, for almost all new work. Opus 5 roughly doubles Opus 4.8's software-engineering score at the same price, with more current knowledge and better alignment. The only reason to stay on 4.8 is a pinned version for reproducibility.
What is Claude Opus 5's knowledge cutoff?
Claude Opus 5 has a knowledge cutoff of May 2026, the most recent of any Claude model, meaning it has more up-to-date training data than earlier releases.
The verdict
Claude Opus 5 is the clearest upgrade Anthropic has shipped in a while: roughly double the software-engineering capability of Opus 4.8, a 3x lead on abstract reasoning, better alignment, and no price increase. For existing Claude users, switching is a config change with immediate upside. For teams evaluating frontier models, Opus 5 makes a strong case as the best value in the top tier — near-Fable-5 intelligence at Opus pricing.
Our recommendation: make Opus 5 your default for hard coding and agentic work, keep Sonnet 5 for high-volume lighter tasks, and re-run your own eval set to confirm the token-efficiency gains on your workloads. If you're still deciding between the major options, start with our comparison of the best AI coding assistants in 2026 — then come back and switch the model string. The upgrade pays for itself in the first hard task it solves that 4.8 couldn't.
Top comments (0)