Claude Opus 5: Anthropic halves frontier AI pricing — what it changes for your SME’s AI agents

On July 24, 2026, Anthropic released Claude Opus 5, priced at $5 per million input tokens and $25 per million output tokens — unchanged from the previous Opus 4.8, but half the price of its more capable sibling Fable 5 ($10/$50 per million tokens), while landing within a point or two of Fable 5 on coding and agentic benchmarks such as SWE-bench. VentureBeat and CNBC both frame it the same way: near-frontier intelligence at a mid-tier price. For any SME already running AI agents in production, or that shelved a project because the per-request cost made the math too tight, this is the third time in 2026 that the price of the underlying model has moved enough to be worth a second look.

What actually changed on July 24

• →Pricing stays at $5/$25 per million input/output tokens — the same as Opus 4.8 — but Opus 5 now performs close to Fable 5 on real coding and reasoning tasks, at half Fable 5’s price.
• →A new “effort” dial (low, medium, high) lets you choose, task by task, how much reasoning the model spends before answering — a cheap, fast pass for simple requests, a slower and pricier one for cases that need it.
• →The headline savings are not the real ones: independent developer benchmarking found Opus 5 uses more tokens per task than Fable 5, so the actual bill drops by roughly 20%, not 50% — worth verifying on your own usage before rebudgeting.
• →Opus 5 ships with no mandatory data retention on the API, a detail enterprise buyers with confidentiality or compliance requirements will care about more than the benchmark scores.

Why a cost dial, not just a cheaper model

The effort dial is not a marketing gimmick — it reflects a documented trade-off in how these models work. A 2024 study by Charlie Snell, Jaehoon Lee, Kelvin Xu and Aviral Kumar, Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters, showed that letting a model “think longer” at inference time — rather than simply using a bigger model — can match or beat a much larger model’s accuracy on hard problems, for a fraction of the cost, provided the extra compute is spent only where the task actually needs it. That is precisely what an effort dial lets a business do deliberately: cheap and fast for routine requests, expensive and careful for the ones where a mistake is costly.

What it means for your SME

Every agent we put into production already makes this choice per use case, not per contract: an inbox triage agent sorting incoming emails runs on a fast, low-effort pass because a wrong classification is cheap to correct, while a regulatory watch agent summarizing a legal text, or a customer support agent handling a complaint, justifies a slower and more careful one because getting it wrong costs more than the extra tokens. If you shelved an AI agent project earlier this year because the projected per-request cost made the return on investment too thin, the arithmetic is worth redoing now — on your own volumes, not on the announcement’s headline number. And if your agents already handle anything sensitive — contracts, customer records, financial data — the no-mandatory-retention detail is one more argument for reviewing which model tier each of your agents actually needs.

Before rebudgeting: four concrete checks

• →Pull your current monthly token volume per agent (input and output separately) before comparing Opus 5’s per-million-token price against what you pay today — the sticker price alone tells you nothing about your real bill.
• →Test the same task at low, medium and high effort on a sample of real requests, and measure both the answer quality and the token count each level actually consumes.
• →For any agent touching confidential or regulated data, check whether Opus 5’s no-mandatory-retention API terms change your data processing agreement or your GDPR documentation.
• →Do not extend a stalled project on the announcement alone — rerun the ROI calculation on your actual volumes before deciding whether the math has genuinely changed.

The number to remember from this launch is not $5 or $25 per million tokens — it is that the cost of running an AI agent keeps shifting, in both directions, several times a year now. An SME that revisits its agents’ economics each time a new model ships stays ahead of the ones that set a budget once and never look again.