Anthropic Opus 5: Same Price, Lower Cost Per Task, No Capability Leap

Anthropic released Opus 5 on July 25, 2026, positioning it as a cost-efficient alternative to its flagship Fable model rather than a capability step-change. Benchmarks including Frontier-Bench and DeepSWE show Opus 5 performing at or slightly above Fable on coding tasks, and ahead of OpenAI's GPT-5.6-Sol, while pricing stays flat at $5 per million input tokens and $25 per million output tokens. The pitch is near-Fable quality at roughly half the cost, with deliberate limits on cybersecurity exploitation training baked in.

What happened

Detail	Fact
Model	Anthropic Opus 5
Input price	$5 per million tokens
Output price	$25 per million tokens
Compared to Fable cost	Approximately half the cost
Coding benchmarks	At or slightly above Fable; ahead of GPT-5.6-Sol
Cybersecurity exploitation	Substantially behind Mythos 5 and Fable by design
Key competitor	Kimi K3 at $15 per million output tokens

Anthropic’s Opus line has become a go-to choice among developers for coding and software work. Opus 5 continues that trend but does not repeat the kind of jump seen with Opus 4.5. According to Anthropic’s own benchmark data, performance improvements are iterative: better than Opus 4.8 and GPT-5.6-Sol across most task categories, roughly level with Fable on coding, but not dramatically ahead of anything released in the past few months.

The pricing is unchanged from the previous version, which means users get more performance per dollar spent without paying more. Compared to Fable, which handles more advanced tasks, Opus 5 delivers similar coding results at about half the price.

What Anthropic left out on purpose

Opus 5 was not given cutting-edge training on cybersecurity exploitation. Anthropic says the model can identify vulnerabilities reasonably well, but it is “substantially behind Mythos 5 on the exploitation of those vulnerabilities.” That is a deliberate training choice, not a technical limitation.

One consequence: Opus 5 does not carry the same content and data policies that came with Fable, including Fable’s 30-day data retention window for incident review. Teams working in security-sensitive contexts should note that gap before swapping models.

Why does the cost angle matter more than the benchmarks?

The developer conversation right now is largely about spend. Frontier models are expensive, and the gap between them and smaller or open-weight alternatives is narrowing fast. Companies like Cursor and Meta are already building model routers, systems that automatically pick a smaller or cheaper model when the task does not require top-tier capability. The goal is to avoid paying Fable-level prices for tasks that a lighter model can handle.

Opus 5 fits into that picture as a middle tier: strong enough for serious coding work, priced low enough to use more freely than Fable. But competition is tightening. The recently announced Chinese open-weight model Kimi K3 comes in at $15 per million output tokens at comparable performance levels, undercutting Opus 5 by $10 per million on output. That is a gap Anthropic will need to answer.

For a closer look at how Chinese open-source models are closing the gap with US frontier labs, see our earlier coverage of Chinese open-weight AI models challenging Silicon Valley.

Our take

Opus 5 is a sensible release, not an exciting one. Anthropic is threading a real needle: keep the Opus line competitive on price while Fable holds the prestige position. That is a reasonable product strategy. The problem is that “more performance for the same money” is now the baseline expectation, not a selling point.

The Kimi K3 number is the one to watch. If open-weight models at $15 per million output tokens keep improving on coding tasks, the rationale for paying $25 weakens fast. Model routers make this even sharper: a business that routes 70% of its coding prompts to a cheaper model and reserves Opus 5 for the hard stuff will spend far less than one running everything through a single frontier model.

If you are using Claude for production workflows and haven’t reviewed your model selection logic recently, now is a good time. Teams we work with through our AI integration service are already building tiered model setups to control costs without sacrificing output quality. The tooling to do this well exists today.

What to do about it

1. Audit your current model spend: break out which tasks are consuming the most tokens and whether they actually need frontier-level performance.
2. Test Opus 5 against your Fable or GPT-5.6-Sol workflows on coding tasks specifically, and measure output quality vs. cost difference.
3. If you handle cybersecurity tooling or vulnerability research, keep Fable or Mythos 5 for exploitation tasks. Opus 5 is not trained for that depth.
4. Evaluate model routing options. Cursor, LiteLLM, and similar tools let you set rules for when to use a lighter model vs. a frontier one.
5. Keep an eye on Kimi K3 and other open-weight releases. The $15 output token price is a credible benchmark for what capable coding models should cost.

The practical takeaway: treat Opus 5 as a cost optimization lever, not a capability upgrade, and build your model stack accordingly.

Source: Ars Technica · AI