Claude Opus 5 Is Anthropic’s New Daily Driver: Frontier Intelligence at Half the Price

juillet 26, 2026 admin

By OFOKI TECH | July 27, 2026

Anthropic released Claude Opus 5 on July 24, 2026. The announcement, published on the company’s official blog, positions the model as the new default for serious coding and knowledge work — not as an incremental update to Opus 4.8, but as a tier that approaches the frontier intelligence of Claude Fable 5 at half the cost. The pricing is identical to Opus 4.8: $5 per million input tokens and $25 per million output tokens. The context window is 1 million tokens, with a maximum output of 128,000 tokens per request.

The release comes two months after Opus 4.8 and one month after the Claude 5 family (Fable 5, Mythos 5, and Sonnet 5) launched in June. It leaves only the lightweight Haiku model without a 5-series upgrade.

What Opus 5 Actually Delivers

Anthropic’s official framing is direct: Opus 5 is designed to be used every day. It is the new default model on Claude Max and the strongest model available on Claude Pro. The company emphasizes that the model is « thoughtful and proactive, » with a specific focus on self-verification — it checks its own work, catches its own errors, and iterates until it succeeds.

On Frontier-Bench v0.1, an agentic terminal coding benchmark that measures real software engineering tasks across multiple files, Opus 5 scores 43.3%. That is more than double Opus 4.8’s 18.7% and nearly 10 points above Fable 5’s 33.7%. On CursorBench 3.2, at maximum effort, the model lands within 0.5% of Fable 5’s peak score, but at half the cost per task.

The standout result, however, is ARC-AGI 3. This benchmark, created by François Chollet, drops AI agents into interactive environments with no instructions, no rules, and no stated goals. The agent must figure out everything through trial and error. Opus 5 scores 30.2%. That is roughly four times the previous best (GPT-5.6 Sol at 7.8%) and twenty times Opus 4.8’s 1.5%. Anthropic notes that humans still clear 100% of the environments, so the benchmark is far from saturated, but the jump resets expectations for the field.

Other published results include:

Zapier AutomationBench: Opus 5’s pass rate is approximately 1.5× the next-best model for the same cost per task.
OSWorld 2.0: The model outperforms every other model at any given cost, surpassing Fable 5’s best result at just over a third of the cost.
Life sciences evaluations: Improvements over Opus 4.8 on structural biology, organic chemistry, and bioinformatics, with the largest gains on protein-related tasks.

The Knowledge Cutoff and the Effort Dial

Opus 5 carries a May 2026 knowledge cutoff, four months fresher than Fable 5 and Opus 4.8. For developers working with recent tooling releases, provider APIs, or Kubernetes deprecations, this matters.

The model introduces an effort parameter that users can set to low, medium, high, or max. At medium effort, Opus 5 matches Sonnet 5’s best score on SWE-bench Verified while using 76% fewer output tokens. At high effort — the default — it exceeds Sonnet 5 by 4.3 percentage points while using 48% fewer tokens. At max effort, it reaches its ceiling on the most complex tasks. The dial allows users to optimize for speed and cost on simple tasks, then crank the reasoning budget up when the problem demands it.

Fast Mode and the API

Anthropic is also offering a Fast mode that runs at roughly 2.5× the default speed for twice the base price. This is currently a research preview, waitlist-gated on the API, and reachable through usage credits in Claude Code. Anthropic warns users not to build latency budgets around it before their account has access.

On the API, the model ID is claude-opus-5. The company has rolled out two beta features alongside the launch:

Mid-conversation tool changes: Developers can now change which tools Claude can use within a conversation without invalidating the prompt cache.
Automatic fallbacks: API requests flagged by safety classifiers on Opus 5 or Fable 5 can automatically route to another model instead of returning an error.

Data Retention and Safety

Unlike Fable 5 and Mythos 5, Opus 5 is not subject to the 30-day data retention policy that Anthropic applies to its frontier models for safety monitoring. For privacy-conscious users and enterprises, this is a meaningful distinction. Anthropic confirms that Opus 5 has no data retention requirements for general access.

Safety classifiers still exist, particularly around cybersecurity tasks like exploit generation and penetration testing. Opus 5 is permitted to scan source code for vulnerabilities but blocked from binary-based scanning. Anthropic states that these classifiers trigger 85% less often than on Fable 5. Flagged requests fall back to Opus 4.8 automatically if the user has enabled automatic fallbacks.

What Anthropic Has Not Clarified

The company has not published independent third-party verification of its benchmark figures. All scores cited above come from Anthropic’s own evaluation runs. While the methodology is described in the announcement, independent labs have not yet confirmed the Frontier-Bench v0.1 or ARC-AGI 3 figures.

The Fast mode remains in research preview with no public timeline for general availability. Early adopters should not architect latency-sensitive applications around it until access is guaranteed.

Anthropic has also not announced a release date for Haiku 5, leaving the budget tier of the Claude 5 family incomplete.

Practical Advice for Developers

Start with Sonnet 5, escalate to Opus 5 on evidence. Sonnet 5 is priced at $3/$15 per million tokens ($2/$10 through August 31, 2026) and handles the majority of production workloads. Route traffic to Opus 5 only when instruction-following precision, long-context work, or complex agentic coding measurably reduces rework.

Use the effort dial. Set low effort for classification, routing, and simple extraction. Set medium for standard coding. Reserve high and max for debugging, architecture decisions, and multi-file refactors. Anthropic’s own data suggests medium effort is the sweet spot for cost-efficiency.

Enable prompt caching. For repeated system prompts, tool definitions, or document prefixes, caching cuts input costs by up to 90%. Cache writes cost 1.25× standard input, so the savings materialize after the first reuse.

Use the Batch API for asynchronous work. It takes a flat 50% off both input and output. Combined with caching, a batched Sonnet 5 workload can approach Haiku pricing while retaining Sonnet quality.

Do not build around Fast mode yet. It is waitlist-gated and priced at 2× base rate. Wait for general availability before committing latency budgets.

Bottom Line

Claude Opus 5 is Anthropic’s attempt to make frontier intelligence a daily tool rather than a specialty service. The pricing is unchanged from Opus 4.8, the benchmarks show double-digit gains over its predecessor, and the effort dial gives users granular control over the speed-versus-quality tradeoff. The absence of data retention requirements and the lighter safety classifiers make it more palatable for enterprise adoption than Fable 5.

But the benchmark figures are still self-reported, the Fast mode is not broadly available, and the $5/$25 price point remains a premium over Sonnet 5’s $3/$15. Whether Opus 5 becomes the default for developers will depend on whether its real-world performance justifies the 1.67× cost multiplier over Sonnet in production environments. Anthropic has laid the numbers on the table. Independent verification will decide if they hold up.

Sources and further reading

Laisser un commentaire Annuler la réponse