Quick answer. Claude Opus 5, launched July 24, 2026, is Anthropic's new near-frontier model: it reaches roughly Claude Fable 5–level intelligence at half the price ($5 per million input tokens, $25 output), adds a low/medium/high effort toggle so you can trade cost for capability per request, and sets new state-of-the-art scores on agentic-coding and knowledge-work benchmarks. It ships with a 1M-token context window, is the default model on Claude Max, and is available in the API as claude-opus-5.
Anthropic shipped its fourth model in under two months on July 24, 2026. After Mythos 5, Fable 5, and Sonnet 5 all landed in June, Claude Opus 5 arrives as the model most teams will actually run every day: it comes within touching distance of the Fable 5 frontier on real work, but costs half as much and adds a control that none of its siblings shipped with — a dial for how hard the model should think before it answers.
This guide pulls apart what's new, what the benchmarks actually say, what it costs once you account for the effort dial, and when you should reach for Opus 5 versus Fable 5, Sonnet 5, or the model you're running today. Numbers below come from Anthropic's launch materials and independent benchmark trackers; where a figure is Anthropic's own framing rather than a third-party score, we say so.
What is Claude Opus 5?
Claude Opus 5 is Anthropic's new general-purpose flagship-adjacent model — described by the company as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." It is positioned as the everyday workhorse for coding, agentic tasks, and knowledge work, while Anthropic still recommends the larger Fable 5 for the most advanced jobs, including agents that run autonomously for days at a time.
The headline specs:
- API model ID:
claude-opus-5 - Context window: 1M tokens (the default and the maximum)
- Max output: 128K tokens standard; up to 300K via the Message Batches API with a beta header
- Extended thinking: on by default
- Effort control: low / medium / high per request
- Pricing: $5 per million input tokens, $25 per million output tokens (standard mode)
- Availability: Claude.ai, Claude Code, Claude Cowork, the Claude API and Platform; default on Claude Max, strongest model on Claude Pro
What makes Opus 5 interesting isn't a single benchmark — it's the price-performance shape. Anthropic's argument is that for the vast majority of production work, you no longer need to pay frontier prices to get near-frontier results.
What's new versus Claude Opus 4.8?
Opus 4.8 shipped on May 28, 2026. Two months later Opus 5 is a clear generational step, not a point release. The changes that matter in practice:
It verifies its own work
The behavioral headline is self-correction. Anthropic highlights that Opus 5 "excels at verifying its work and iterating carefully until it succeeds," and cites an example where the model wrote its own computer-vision pipeline from an incomplete prompt — filling gaps, testing, and fixing its own mistakes without a human nudging it back on track. In agent loops that means fewer rounds of "no, that's wrong, try again."
The effort toggle
New in this release: you can tell Opus 5 to spend low, medium, or high effort on a task. Low burns fewer thinking tokens and returns faster and cheaper; high lets the model reason as long as it needs for hard problems. This is the "balance cost and capability" lever Anthropic built specifically for teams worried about runaway inference bills. We cover the cost math below.
A big jump in agentic coding
On Frontier-Bench v0.1, Anthropic's agentic-coding evaluation, Opus 5 "more than doubles" Opus 4.8's pass rate — while costing less per task. That's the single largest capability delta between the two models.
Better science, safer defaults
Opus 5 improves on Opus 4.8 across every life-sciences evaluation Anthropic ran, with 10-plus-point gains in areas like organic chemistry; the company calls it "its most capable generally available model for scientific research." On the safety side, it has stronger cybersecurity guardrails than 4.8 (it can examine source code but won't scan compiled binaries for vulnerabilities), and Anthropic's automated behavioral audit scores it as its most-aligned model to date.
How does Opus 5 do on benchmarks?
The through-line across Anthropic's launch numbers is that Opus 5 wins on capability-per-dollar, and on several evals wins outright. Here are the hard scores that have public figures:
| Benchmark | Claude Opus 5 | Claude Fable 5 | GPT-5.6 Sol |
|---|---|---|---|
| Frontier-Bench v0.1 (agentic coding, % passed) | 43.3% | 33.7% | 34.4% |
| ARC-AGI-3 (novel problem-solving) | 30.2% | — | 7.8% |
| GDPval-AA v2 (knowledge work, Elo) | 1,861 | 1,747 | 1,736 |
| SWE-bench Pro | 79.2% | 80.0% | — |
A few things stand out:
- Agentic coding is the strongest result. Opus 5's 43.3% on Frontier-Bench beats both Fable 5 and OpenAI's GPT-5.6 Sol outright — a rare case of the cheaper model topping the flagship.
- ARC-AGI-3 is a blowout. At 30.2% versus 7.8% for the next-best model, Opus 5 is roughly three times better at the kind of novel, no-memorization puzzles that stress genuine reasoning.
- Knowledge work leads too. On GDPval-AA v2, a human-graded Elo of real economic tasks, Opus 5 sits above Fable 5 and GPT-5.6 Sol.
- SWE-bench Pro is the one place it's third, trailing Mythos 5 (80.3%) and Fable 5 (80.0%) by under a point — still far ahead of Opus 4.8's 69.2%.
Anthropic also shared framed (non-raw) results worth knowing: on CursorBench 3.2 at max effort, Opus 5 lands within 0.5 points of Fable 5's peak at roughly half the cost per task; on Zapier AutomationBench it completed a full end-to-end churn-prevention workflow and runs about 1.5× the next-closest model at matching cost; and on OSWorld 2.0 (computer-use) it "outperforms rivals at every price point," beating Fable 5's peak using about one-third the budget.
How the effort toggle changes the cost math
The effort dial is the feature that makes the pricing genuinely different from "Opus 4.8 but cheaper." Because Opus 5 has extended thinking on by default, the tokens it spends reasoning are billed as output. Setting effort to low caps that reasoning, so a routine task — a lint fix, a short summary, a well-scoped function — costs a fraction of what the same call costs at high effort. Reserve high for genuinely hard problems where a wrong answer is expensive.
In practice, that means the sticker price understates how cheap Opus 5 can be. A team that routes 80% of its traffic through low/medium effort and only escalates the hard 20% to high will pay far less than the per-token rate suggests — and still get frontier-class reasoning where it matters.
What does Claude Opus 5 cost?
Opus 5 keeps the same headline rate as Opus 4.8, which is the point: you get a generational capability jump at no price increase, and at half the input price of Fable 5.
| Model / mode | Input ($/M tokens) | Output ($/M tokens) |
|---|---|---|
| Opus 5 — standard | $5 | $25 |
| Opus 5 — fast mode | $10 | $50 |
| Opus 4.8 | $5 | $25 |
| Fable 5 (frontier) | $10 | — |
Fast mode costs 2× the standard rate and runs about 2.5× faster — useful for latency-sensitive, interactive work in Claude Code or the Platform. There's one more cost-adjacent detail worth flagging for privacy-sensitive teams: unlike Fable 5, Opus 5 is not subject to the 30-day data-retention policy, and Anthropic expects safety classifiers to trigger about 85% less often than on Fable 5, so you should hit far fewer spurious refusals in production.
Where can you use Opus 5?
Opus 5 is available everywhere the Claude 5 family lives:
- Claude.ai — the new default model on Claude Max, and the strongest model offered on Claude Pro.
- Claude Code — as the model behind Anthropic's terminal agent, with fast mode available.
- Claude Cowork — Anthropic's collaborative workspace.
- Claude API / Platform — as
claude-opus-5, with the effort parameter and fast mode exposed to developers.
There's also a new beta safeguard called Automatic Fallbacks: when a safety classifier trips, instead of returning an error the Platform can route the request to a less-powerful model and still return a usable response — handy for keeping automated pipelines from hard-failing on false positives.
Opus 5 vs Fable 5: which should you use?
This is the decision most teams are actually making today. The short version:
- Reach for Opus 5 for the overwhelming majority of coding, agentic, and knowledge work. It matches or beats Fable 5 on most public benchmarks, costs half as much on input, and carries fewer data-retention and refusal frictions.
- Reach for Fable 5 when a task is genuinely at the edge of what's possible — long-horizon agents that run autonomously for days, or the hardest research problems where you want the absolute top of Anthropic's lineup and cost is secondary. Anthropic itself still recommends Fable 5 there.
- Reach for Mythos 5 specifically for frontier cybersecurity exploitation and the most demanding biology research, where it still leads Opus 5.
For a full breakdown of the frontier tier, see our Claude Fable 5 launch guide and the Anthropic Mythos guide.
Opus 5 vs GPT-5.6 Sol: the cross-vendor picture
Against OpenAI's GPT-5.6 Sol, Opus 5 leads on the evals with public numbers: 43.3% vs 34.4% on Frontier-Bench agentic coding, 30.2% vs 7.8% on ARC-AGI-3, and 1,861 vs 1,736 on GDPval-AA v2. The gap is widest on novel reasoning (ARC-AGI-3) and agentic coding — the two areas most teams care about for autonomous developer workflows. If your stack is coding- or agent-heavy, Opus 5 is the stronger and cheaper pick; if you're already deep in the OpenAI ecosystem, the head-to-head is close enough on knowledge work that switching cost matters more than raw score. Our GPT-5.6 vs Claude Fable 5 comparison has the deeper cross-vendor dive.
Opus 5 vs Sonnet 5 and Opus 4.8
Inside the Claude lineup, the ladder is now clear:
- Sonnet 5 remains the fast, cheap workhorse for high-volume, well-scoped tasks where you don't need frontier reasoning. See the Claude Sonnet 5 launch guide.
- Opus 5 is the new near-frontier default — reach for it when a task needs real reasoning, self-verification, or long-context work, and you want that without Fable 5 prices.
- Opus 4.8 is now superseded: same price as Opus 5, meaningfully weaker on coding and reasoning. There's little reason to start new work on 4.8. If you're on it today, see the Opus 4.8 guide for the migration context.
Should you switch to Claude Opus 5?
For most teams the answer is yes, and the switch is low-risk:
- If you're on Opus 4.8: upgrade. Same price, generational jump, drop-in model ID change.
- If you're on Fable 5 for everything: route your routine traffic to Opus 5 and keep Fable 5 for the genuinely hard 10–20%. You'll cut cost sharply with little quality loss.
- If you're on a competitor: Opus 5's agentic-coding and reasoning lead, plus the effort toggle and no 30-day retention, make it worth a serious evaluation — especially for developer-tool and agent workloads.
As always, run your own eval on your own tasks before rewiring production. But on launch-day evidence, Opus 5 is the most sensible default in Anthropic's lineup for the first time in a while.
Related guide
New to running frontier models in your dev workflow? Our AI Coding Agents complete guide covers how to wire Opus 5, Claude Code, Cursor, and Codex into a real engineering team — and the Claude Opus developer guide goes deep on the API, tool use, and agent patterns.
FAQ
When was Claude Opus 5 released?
Anthropic launched Claude Opus 5 on July 24, 2026 — its fourth model in under two months, following Mythos 5, Fable 5, and Sonnet 5 in June.
How much does Claude Opus 5 cost?
Standard mode is $5 per million input tokens and $25 per million output tokens — the same rate as Opus 4.8 and half of Fable 5's input price. Fast mode doubles the rate ($10/$50) and runs about 2.5× faster.
What is the effort toggle in Opus 5?
It's a per-request setting — low, medium, or high — that controls how much reasoning effort the model spends. Low is faster and cheaper for routine work; high lets the model think longer on hard problems. It's Anthropic's lever for balancing cost against capability.
Is Opus 5 better than Fable 5?
On most public benchmarks Opus 5 matches or beats Fable 5 at half the input price, and it wins outright on agentic coding (Frontier-Bench) and novel reasoning (ARC-AGI-3). Anthropic still recommends Fable 5 for the most advanced, long-horizon autonomous work, and Mythos 5 leads on frontier cybersecurity and biology.
What is the context window on Claude Opus 5?
Opus 5 has a 1M-token context window (both the default and the maximum), with 128K max output tokens in standard requests and up to 300K via the Message Batches API with a beta header.
What is the API model ID for Claude Opus 5?
It's claude-opus-5. Extended thinking is on by default, and the effort parameter plus fast mode are exposed to developers on the Claude API and Platform.
Building with frontier models but short on senior engineers?
Codersera matches you with vetted, remote-ready developers who ship with AI coding agents like Claude Opus 5 in their daily workflow. Extend your team in days, not months — risk-free.