Anthropic released Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output — unchanged from Opus 4.8. The pitch: performance within 0.5% of the frontier-tier Fable 5 at half the cost, per the company's own benchmarks. For operators, the headline is not the leaderboard. It is the repricing. Capability that cost frontier rates this spring now costs half, and any routing table or budget model built before last week is due for a re-run.
Key takeaways
- Same price, higher claimed capability: Opus 5 costs $5 per million input tokens and $25 per million output — identical to Opus 4.8 — and Anthropic claims results within 0.5% of Fable 5 on CursorBench 3.2 at half the cost.
- Broad availability from day one: Claude.ai, the API, Claude Code, and Claude Cowork, with no special data-retention requirements for general access.
- The frontier is repricing fast: Axios counts four Claude 5-generation releases in under two months. Model spend math now has a shelf life measured in weeks.
- Benchmarks are vendor claims: the number that matters for your systems is cost per passed acceptance test, measured on your own workloads.
- Model-agnostic design pays out today: systems built behind acceptance criteria can trial the swap in days; systems hard-wired to one model renegotiate their architecture first.
What Anthropic shipped
The release is broad from the start. Claude Opus 5 is available on Claude.ai, through the API under the model ID claude-opus-5, and in Claude Code and Claude Cowork, with no special data-retention requirements for general access. It is the new default model on Claude Max. A fast mode is offered at 2x the base price and, per Anthropic, runs about 2.5x faster.
The performance figures are Anthropic's. The company claims Opus 5 lands within 0.5% of Fable 5 on CursorBench 3.2 at half the cost, surpasses Fable 5 on OSWorld 2.0 at one-third the cost, and scores roughly 3x the next-best model on ARC-AGI 3.
There is a safety story too. Anthropic describes Opus 5 as its "most aligned model to date" per behavioral audit, says the model was intentionally not trained on cyber exploitation, and reports its cyber classifiers intervening about 85% less often than with Fable 5. If your security team gates model changes, those are the claims to put in front of them.
The frontier is repricing itself
Axios noted that this is Anthropic's fourth Claude 5-generation model release in under two months, and that Opus 5 debuted at the top of the Artificial Analysis leaderboard.
The cadence matters more than the ranking. Leaderboard positions change hands monthly. What changed here is the price of a given level of capability: work that justified frontier rates in the spring can now be quoted at half — if the vendor's parity claims hold on your tasks. A per-task cost model locked in during Q2 is already stale.
Price per token is the vendor's number. Price per passed acceptance test is yours.
Cost per passed acceptance test, not leaderboard position
Every performance figure above was measured on a public benchmark. Your workload is not a public benchmark. The number an operator needs is different: total spend on a workload, divided by the acceptance tests it passes.
That metric is cheap to compute if the tests exist. Run the incumbent model and Opus 5 through the same acceptance suite. Count passes. Divide by cost. The cheaper passing model wins, whatever the leaderboard says.
This is also where architecture pays out — or bills you. Systems designed model-agnostically, with the model behind an interface and behavior pinned by acceptance criteria rather than one model's quirks, can trial claude-opus-5 behind unchanged criteria in days. Systems with prompts hand-tuned to a single model pay a migration tax on every release. At the current cadence, that tax comes due every few weeks.
What to do this week
- Re-run the routing math. If you pay frontier rates for agentic workloads, Anthropic's claimed parity at half the cost is worth a same-week evaluation on your own acceptance suite.
- Treat the Opus 4.8 upgrade as near-free — after regression. Same price, higher claimed capability. Run the suite before flipping defaults; near-free is not zero-risk.
- Price fast mode against latency-sensitive flows. 2x cost for roughly 2.5x speed, per the company, reads differently for customer-facing work than for overnight batch jobs.
- Log swap friction as a finding. If evaluating a new model means rewriting prompts and hand-checking outputs, the architecture is the problem. Fix that before the next release lands — there will be one.
We have argued before that acceptance criteria, not demos, should decide what ships — that piece is the companion to this one, and our GPT-5.6 briefing earlier this month covered the other side of the same fast-moving market. If your systems cannot take advantage of capability that got twice as cheap in a quarter, that is a design problem — see how we build custom AI applications, or book a consult.