Claude Opus 5: near-frontier intelligence at half the price — and it polices its own work

Why the price reset matters globally

For two years, model selection at most companies followed a simple rule: run the affordable model in production, reserve the frontier option for the work important enough to justify the bill. Opus 5 collapses that separation. At $5 / $25 per million tokens it sits in the old Opus tier, yet on the partner benchmarks Anthropic published it tracks the $10 / $50 Fable 5 closely enough to force a question every buyer now has to answer: how often is the remaining capability gap actually worth paying 2× for?

The question is not academic. Per the official Anthropic pricing page, Opus 5 is now the default model on Claude Max and the strongest model on Claude Pro. That changes default routing inside coding assistants and agent frameworks — the model most developers reach for on a Tuesday afternoon has moved up a tier without moving up in price.

For global buyers holding budgets in USD, EUR or GBP the math is straightforward. Put it against the current Chinese frontier open model, Kimi K3 from Moonshot (1 $/3 $ per million tokens, 2.8T parameters, open weights), and Opus 5 is roughly five times more expensive per raw token. But Kimi K3 does not yet match Anthropic's numbers on long-horizon agentic coding or on benchmarks like OSWorld 2.0. For many production teams the honest comparison is no longer "frontier versus cheap" — it is "near-frontier at Opus price, frontier at Fable price, or self-hosted at cost of GPUs and ops."

The four numbers that explain the launch

Anthropic brought a wall of benchmarks to the release. Four results carry most of the weight:

• Frontier-Bench v0.1 — Opus 5 surpasses every tested model, and Anthropic reports it more than doubles Opus 4.8 at a lower task cost.
• CursorBench 3.2 — at max effort Opus 5 lands within 0.5% of Fable 5's peak while costing half as much per completed task.
• ARC-AGI 3 — a novel-problem-solving evaluation. Anthropic reports Opus 5 at 30.2%, versus 7.8% for GPT-5.6 Sol and 1.5% for Opus 4.8.
• OSWorld 2.0 — a computer-use benchmark. Opus 5 reportedly beats Fable 5's best result at just over one-third of the cost.

The clean sweep is not entire. Anthropic's own System Card states plainly that Opus 5 stays behind Mythos 5 in biology research and offensive cybersecurity, and that GPT-5.6 Sol still leads on DeepSWE. The launch claim is narrower than the headlines imply: Opus 5 did not kill Fable 5. It made Fable 5 harder to justify for everyday work.

The behaviour that changes cost in production

The single most consequential trait in Opus 5 is not a context number — it is a behaviour. Anthropic's prompting guide explicitly tells developers to remove legacy instructions like "include a final verification step" or "use a subagent to double check" because Opus 5 already does this and adding more verification just burns tokens repeating work.

The Frontier-Bench FreeCAD case is illustrative. Given a technical drawing it could not view, Opus 5 wrote its own computer-vision pipeline, extracted geometry from raw pixels, rebuilt the part in FreeCAD, and reproduced the result. No competing model with the same constrained setup solved it in five attempts. The pattern recurs across early-access reports: when a trading engineer asked Opus 5 to build a market-data feed with no live stream to validate against, the model built its own test harness before returning the parser.

This is where reliability becomes an economic feature. The real cost of an AI task is not the first generation — it is the engineer who reviews it, the correction prompt, the rerun, the hidden edge case, the rollback. If Opus 5 catches more of those failures before handoff, the biggest saving comes from fewer failed handoffs rather than cheaper tokens.

Five effort levels turn one model into a cost curve

Opus 5 ships with five effort settings — low → medium → high → xhigh → max — and thinking is on by default. The 1-million-token context window and up to 128,000 output tokens give it the headroom for repository-scale work. Practical guidance from Anthropic is to pick the lowest effort level that still produces a trusted result rather than maxing intelligence on every request. There is also a Fast mode running roughly 2.5× the default speed at double the token price ($10 / $50) — at which point Opus 5 costs the same per token as standard Fable 5. Speed is another dial, not a free upgrade.

EU and Czech angle: data residency, GDPR, AI Act

For European buyers, model intelligence is one question; where the data goes is the other. Opus 5 is available on Claude API, Amazon Bedrock (global and regional endpoints), Google Cloud Vertex AI (global, multi-region and regional endpoints) and the new Claude Platform on AWS. The regional and multi-region endpoint options let a Czech bank, a hospital or a public-sector buyer keep inference inside the EU — directly relevant for GDPR data residency and for the data-sovereignty clauses public tenders increasingly require.

Claude API itself carries no explicit data retention requirement on Anthropic's side for general access, but teams running sensitive workloads should still audit backup, logging and subprocessor terms with their cloud provider. Under the EU AI Act, Opus 5 falls in the general-purpose AI category — what makes the use compliant is chiefly the deployer, not the lab in San Francisco. That moves the practical compliance burden for European teams away from the model and toward prompts, logging and the governance layer around the agent.

Three ways to destroy the savings

Half the token price does not guarantee half the bill. Three mistakes quietly erase it:

1. Over-verification. Leaving old "double-check everything" prompts in place makes Opus 5 do work twice.
2. Uncontrolled subagents. Opus 5 delegates more readily than earlier Opus models; one task can quietly become five agents reviewing each other, multiplying cost.
3. Vague scope. A proactive model easily expands a narrow request into a redesign. Cost-sensitive work should declare the objective, allowed files, protected systems, stopping condition and maximum number of subagents.

The right frame: the best model is not the one with the lowest token price. It is the one that reaches a trusted result with the least total work.

The honest part

Most of the strongest launch numbers come from Anthropic or its early-access partners — Cursor, Cognition, Box, Lovable, JetBrains and others. They are not worthless, but they are not the final word. Independent replication is not yet in. Half the token price also does not guarantee half the final invoice: effort level, output length, tool calls, subagents and retries all still move the bill. Opus 5 does not replace Fable 5, and Anthropic has been open about the cases where Fable and Mythos remain ahead. The real verdict will come from developers measuring cost per completed task against their own codebases and workflows, not from a launch-day leaderboard.

Conclusion

Frontier AI did not become cheap. Opus 5 did something more useful: it moved near-frontier capability into a price tier companies can actually afford to keep running. The gap between the best model on the market and the model teams can run in production just narrowed significantly. The token price stayed where it was. The intelligence moved. If Opus 5 consistently catches its own mistakes before the handoff, the biggest saving will not be cheaper tokens — it will be the work that no longer has to be done twice.