Claude Opus 5: Same Price, Much Better Coding

That's worth pausing on, because frontier AI pricing usually moves in one direction: more capability, more money. Opus 5 breaks that pattern on paper, at least for the base price. Real-world cost will still depend on effort settings, tool calls, and retries, as launch coverage notes, and there's a fast mode at twice the price — but for anyone already paying Opus prices, Opus 5 reads as a straightforward upgrade, not a budget decision.

Anthropic positions Opus 5 as its premium model for complex agentic coding, long-horizon work, and enterprise tasks, below Fable 5 but close to its frontier intelligence at half the price. It's the new default for Claude Max subscribers and the strongest model on Claude Pro, with API access via Anthropic, Bedrock, Google Cloud, and Microsoft Foundry.

One caveat: by October 2026, Anthropic had also pointed migrators at Claude Opus 5.5. Opus 5 is therefore not necessarily the latest model when you read this — check Anthropic's current list before committing to a new integration.

Specs at a glance

The specification sheet is deliberately unremarkable, which is the point, as one launch analysis put it. A 1M-token context window, matching Fable 5, Sonnet 5, and the entire Opus 4.x line, as both default and maximum — no smaller variant. A 128k maximum output on the synchronous Messages API, extending to 300k on the Message Batches API with a beta header. Adaptive thinking, with effort levels from low to max, defaulting to high. The full 1M-token window is billed at standard rates, with no premium above 200k input tokens, according to that analysis.

Two details matter for developers. First, the May 2026 knowledge cutoff is four months more recent than the January 2026 cutoff carried by Fable 5, Opus 4.8, and Sonnet 5, which matters for anything involving recent tooling releases, provider APIs, or Kubernetes deprecations. Second, the old extended-thinking mode — thinking type "enabled" — is gone entirely on Opus 5, replaced by adaptive thinking. If your code still calls the old thinking type, it will break, and you need to migrate to the new effort-based API before you switch models.

There's also a fast mode that runs roughly 2.5 times faster at twice the base price ($10 / $50), as a research preview, waitlist-gated on the API, and reachable through usage credits in Claude Code. Don't build a latency budget around it before your account has access, as launch coverage warns.

What Anthropic claims on benchmarks

Anthropic's launch post leans on five evaluations, and none of them are the familiar SWE-bench or Terminal-Bench numbers earlier launches quoted, according to ClockedCode's analysis. Here's what Anthropic reported, with the caveat that these are vendor-reported scores, not independent tests:

• Frontier-Bench v0.1: Opus 5 more than doubles Opus 4.8's score, at a lower cost per task. One guide gives the figure as 43.3% on Frontier-Bench v0.1, versus Opus 4.8's lower score — more than doubling, as Anthropic claimed.
• CursorBench 3.2: At maximum effort, Opus 5 lands within half a percent of Fable 5's peak result, while costing about half as much per task. That's the core value argument: near-frontier coding at Opus price.
• ARC-AGI 3: Three times the score of the next best model, according to Anthropic — the boldest claim in the launch, for novel reasoning tasks that test generalisation, not memorisation.
• OSWorld 2.0 (computer use): Opus 5 reportedly beats Fable 5's best result, at just over a third of Fable 5's cost.
• Zapier AutomationBench: Pass rate about 1.5 times the next best model, at the same spend, for real automation workflows rather than coding tasks specifically.

Independent verification, on the day of release, came from Artificial Analysis: Opus 5 recorded 61 on version 4.1 of its Intelligence Index at maximum effort, first of 190 models, one point ahead of Fable 5 at 60, according to launch analysis. The 61 is a max-effort number — the API and Claude Code default to high effort, which Artificial Analysis scores at 59, level with GPT-5.6 Sol, according to the same analysis. That's still a strong result, but it's a reminder to check which effort level a benchmark used before you compare scores across models.

One independent test, by ComputingForGeeks on 25 July 2026, put Opus 5 head to head against Opus 4.8 on real Terraform, Kubernetes, and Bash work through the Claude API, with real linters as judge, and verified claims against the live API and official docs. That kind of real-world, linter-judged test is more useful than a benchmark score alone, because it measures whether generated infrastructure code actually passes validation, not just whether it looks plausible.

What actually changed for developers

Beyond benchmarks, Opus 5 changes how you call the model, and how it behaves inside Claude Code.

Adaptive thinking by default. On Opus 4.8, adaptive thinking was opt-in. On Opus 5, it's on by default, with effort levels from low to max, defaulting to high. Disabling thinking is capped at effort "high", according to API migration notes — you can't simply turn thinking off and expect the same behaviour as before. For most coding work, that's a benefit: the model reasons more before it answers, and self-verifies its work. For latency-sensitive, simple tasks, it's a cost: you're paying for reasoning you may not need, and you should set effort to low for those tasks.

Self-verification and error recovery. Opus 5 checks its own work and iterates until it succeeds, according to launch coverage, so you spend less time in the back-and-forth loop of "that didn't work, try again." For long-horizon tasks — a multi-file refactor, a Terraform plan, a data pipeline — that self-correction is the difference between a model that starts well and a model that finishes well.

New beta features. Tools can change during a conversation, and safety-triggered requests can fall back automatically to another model, according to launch coverage. Those are beta features, so don't rely on them in production without testing, but they point to where Anthropic is heading: models that adapt their toolset mid-task, rather than being locked to the tools you gave them at the start.

A cyber-category safety classifier. Opus 5 carries a safety classifier that Opus 4.8 never had, according to migration notes. That can change how it acts inside Claude Code, especially for security-related work — it may refuse, or route to a different model, for requests that Opus 4.8 would have attempted. If you work in offensive cybersecurity, Anthropic still credits Fable 5 with leading in that area, according to migration guidance, so reach for Fable 5 specifically for those workloads.

Migration gotchas. The old thinking type is gone, adaptive thinking is default, and the safety classifier is new. If you were already defaulting to an Opus model in Claude Code, switching costs you nothing on price and should read as a straightforward upgrade — just don't assume old API code ports over unchanged, as ClockedCode warns. Test your prompts, your tool definitions, and your effort settings on Opus 5 before you switch production traffic.

Opus 5 versus Sonnet 5 versus Fable 5: which to pick?

Anthropic's model-selection guidance, as summarised in launch coverage, now reads: start with Claude Opus 5 for complex agentic coding and enterprise work, and reach for Claude Fable 5 only for workloads that need the highest available capability.

In practice, the choice is about cost versus capability versus latency:

• Sonnet 5 ($2 / $10 standard, after Anthropic cancelled the scheduled increase to $3 / $15 on 10 August 2026): the clear default for mid-weight agent work, everyday coding, and high-volume tasks where Opus 5's extra reasoning isn't worth 2.5 times the price. If you're running thousands of agent tasks a day, Sonnet 5's price advantage compounds quickly.
• Opus 5 ($5 / $25): for complex agentic coding, long-horizon work, scientific analysis, financial research, legal work, and complex documents — tasks where the model must plan, use tools, verify its own work, and continue until the objective is complete. It's the workhorse for hard problems, at a price that hasn't risen since Opus 4.5.
• Fable 5 ($10 / $50): for the highest available capability, especially offensive cybersecurity work and biology research, where Anthropic still credits Fable 5 with leading. For most other work, Opus 5's near-frontier scores at half the price make Fable 5 hard to justify, unless you need that last few percent of capability.

A practical rule: start with Sonnet 5, upgrade to Opus 5 when a task fails or needs long-horizon reasoning, and upgrade to Fable 5 only when Opus 5 fails on a frontier task.

Safety, alignment, and the system card

Opus 5 shipped with a 194-page system card, according to launch analysis — Anthropic's detailed safety and evaluation report, covering how the model was tested, where it fails, and what mitigations are in place. Automated behavioural audits rated Opus 5 as Anthropic's most-aligned model to date, according to one guide, which also noted its strongest ARC-AGI-3 novel-reasoning and GDPval knowledge-work scores in Anthropic's lineup.

That's encouraging, but it's not a guarantee. A system card is a vendor report, not an independent audit. For enterprise work, still review outputs, especially production code and financial or legal documents, where a plausible error could cause real harm.

The bottom line

Claude Opus 5 is a step change at an unchanged price: a 1M-token context, May 2026 knowledge, adaptive thinking by default, and vendor-reported scores that more than double Opus 4.8 on Frontier-Bench, land within half a percent of Fable 5 on CursorBench at half the price, and triple the next best model on ARC-AGI 3. Independent Artificial Analysis scoring put it first of 190 models at maximum effort on release day, at 61, versus Fable 5 at 60.

For Claude Code users already on Opus, it's a straightforward upgrade — same price, better scores, no reason to stay on the older model once you've confirmed your code handles the API changes. For new users, it's the model to start with for hard coding and enterprise work, with Sonnet 5 as the cheaper default for mid-weight tasks, and Fable 5 reserved for the highest-capability workloads.

Check Anthropic's current model list before you commit, because Opus 5.5 arrived by October 2026, and the frontier moves fast. But Opus 5's lesson endures: a frontier model doesn't have to cost more to be much better — and when it doesn't, the right move is to test it, migrate your API calls, and let it do the hard work at the old price.