Claude Opus 5: the real cost is token consumption, not the sticker price

On July 24, 2026, Anthropic released Claude Opus 5 — the fourth Claude 5 release in under two months, per Axios. The positioning, as described by Fortune and CNBC, is clear enough: capability approaching Fable 5 — Anthropic's most capable model — at half the price. After two analyses covering Z.ai's GLM-5.2 on the low-cost end and Moonshot AI's Kimi K3 on the open-source end, this is the remaining piece: the frontier closed-model tier. And for an engineering team, the most interesting thing about Opus 5 isn't the benchmark table — it's how it consumes tokens.

A note on how reliable the numbers in this post are

We separate two categories of figures here. The first — pricing, context limits, API parameters — comes straight from Anthropic's official documentation and is verifiable right now. The second — the benchmark numbers — reached us only through press coverage and third-party evaluation sites, and we were not able to check them against the primary source at the time of writing. We cite each figure so you can judge for yourself, rather than presenting everything at the same level of certainty.

Benchmarks: leading, but the margin depends on the test

Artificial Analysis scores Opus 5 at 61 on its Intelligence Index — the highest at time of publication, ahead of Fable 5 (60), GPT-5.6 Sol (59), Kimi K3 (57), and Opus 4.8 (56). More notable than the absolute number: the same evaluator records Opus 5 as roughly 26% cheaper per task than Fable 5, meaning the capability gap is narrow while the cost gap is not.

The benchmarks published by Anthropic, as relayed by Vellum and the-decoder, show margins that vary sharply by task type: Frontier-Bench v0.1 (agentic coding) at 43.3% versus 34.4% for GPT-5.6 Sol and 33.7% for Fable 5; ARC-AGI-3 at 30.2% versus 7.8% for GPT-5.6 Sol — an unusually wide gap; OSWorld 2.0 (computer use) at 70.6% versus 66.1% for Fable 5. This is the familiar pattern we described in our post on AI's jagged frontier: a model that dominates one task category can be merely level in another, which makes picking a model on a composite score the easiest way to get it wrong.

What matters more than list price: tokens actually consumed

Opus 5's pricing is unchanged from Opus 4.8: $5 per million input tokens and $25 per million output tokens, against $10/$50 for Fable 5. Cache hits cost just $0.50 per million, and the Batch API halves it to $2.50/$12.50. The 1M-token context carries no long-context premium, and max output is 128K tokens.

But unit price is only half the equation. CodeRabbit measured that on code-review work, Opus 5 reads roughly 50% more and writes roughly 65% more tokens than other frontier models. The cause is by design: adaptive thinking is on by default, and thinking tokens bill as output. Which means a model that is 50% cheaper on the price sheet can still produce an equal or larger invoice if it thinks longer for the same task.

This is why, when we evaluate model cost for a client, we always run a representative task set and measure real token counts rather than multiplying unit price by an estimate. Unit price is what the vendor publishes; token consumption is what determines the invoice.

The effort parameter — the biggest cost lever, and its traps

Opus 5 exposes an effort parameter with five levels (low, medium, high, xhigh, max), defaulting to high. Artificial Analysis scored each level: max 61, xhigh 60, high 59, medium 56, low 51 — meaning dropping from max to medium costs only 5 points while saving a substantial number of tokens. For most repetitive work in a production system, that's a trade worth serious consideration.

Three details in the official documentation tend to surprise teams in real deployments. First, thinking cannot be disabled at xhigh and max — the API returns a 400 if you try. Second, the docs state plainly that lowering effort does not reliably shorten responses; if you want shorter output, ask for it in the prompt rather than dropping effort. Third, changing effort mid-conversation invalidates prompt caching — a detail that can quietly increase costs if your routing layer switches effort per turn.

How KonexForge fits Opus 5 into the model-routing layer

As with GLM-5.2 and Kimi K3, Opus 5 joins the Router layer of KonexForge AI Core as one option, not a replacement. The difference is that the effort parameter enables far more granular routing than before: instead of choosing only "which model," the Router now chooses "which model at which reasoning level." Boilerplate generation and simple classification run at low effort; architectural decisions and schema migrations run high. For high-volume work that doesn't demand maximum accuracy, GLM-5.2 remains a sensible low-cost choice — the principle of not locking into a single AI provider doesn't change just because a stronger model exists.

On integration readiness, Opus 5 landed in GitHub Copilot on release day for Pro+, Max, Business, and Enterprise plans — though administrators must enable the policy first. Several customer results published via VentureBeat point the same direction on token efficiency: Harvey matched Opus 4.8's maximum-reasoning quality using 26% fewer tokens; Box reported an 8% overall improvement and 17% on due-diligence tasks. These are figures published by Anthropic's own customers, so read them as indicative signals rather than independent measurement.

Limitations and risks worth weighing

• **Higher latency by default** — Fast mode is roughly 2.5× faster but costs double ($10/$50), and is currently available only on Anthropic's first-party API; not on Bedrock, Vertex, or the Batch API. For systems that must run through a cloud provider for data-compliance reasons, that constraint needs confirming early.
• **Cost is hard to predict** — the combination of five effort levels, caching, and Fast mode creates several distinct cost profiles that the invoice doesn't separate cleanly, as finout analyzes. You need your own tagging and measurement at the application layer to know where the money goes.
• **Not the strongest at everything** — Fable 5 remains the choice for the hardest category of work, and some sources report GPT-5.6 Sol leading on long-horizon coding. Several reviews also describe Opus 5 as overly cautious on security-related tasks.
• **Safety findings worth tracking** — the system card rates alignment risk as very low and notes the model over-refuses less than its predecessor, but also records higher evaluation awareness. More notably, TechTimes relays UK AISI test results showing the model successfully compromised simulated enterprise networks in 8 of 10 attempts. For infrastructure teams, that's a reminder that stronger agentic capability means a larger attack surface — exactly the zero-trust principle we set out in our post on security in the age of AI.

Conclusion

Claude Opus 5 narrows the capability gap with Anthropic's own most expensive model while holding its previous price — that's real progress. But the engineering lesson worth taking away isn't in the benchmark table: with a generation of models that have adaptive reasoning on by default, price per million tokens is no longer a sufficient basis for comparing cost. What you need to measure is tokens actually consumed on your specific workload, and the effort parameter is the most direct tool for tuning it. Learn more about KonexForge's AI & ML capability, or get in touch if you want to re-evaluate your current model-routing costs based on real token counts rather than list prices.