Claude Opus 5: Near-Fable 5 Quality at Half the Cost

Anthropic shipped Claude Opus 5 on July 24, 2026, at the same $5/$25 per-million-token pricing as Opus 4.8, with a claim that carries real weight: within 0.5% of Claude Fable 5's peak on CursorBench 3.2 at half the cost per task. It is the fourth Claude 5-series model in under two months, after Sonnet 5, Fable 5, and Mythos 5, and it now anchors Claude Max by default. Here is what changed, what the benchmarks show, and who should switch.

What is new in Opus 5

Opus 5 keeps the Opus price point but adds a five-level effort toggle, from low to max, plus a 1M-token context window. The effort setting is the practical headline: the same model can run fast and cheap for routine work, then spend deeply on hard reasoning when you ask. That is a departure from the fixed-effort Opus line and it is why the "half the cost" claim needs the effort caveat: near-Fable-5 results assume the right effort level for the task.

What the benchmarks show

• CursorBench 3.2: within 0.5% of Fable 5's peak, at roughly half the per-task cost
• Frontier-Bench v0.1: leads all models, more than doubling Opus 4.8's score at lower cost per task
• ARC-AGI 3: 3x the score of the next-best model

Anthropic's own model card adds the honest side: Opus 5 scores 2.3 on overall misaligned behavior, the lowest of its recent models, but still trails Mythos 5 on cybersecurity exploitation.

The Fable 5 context

Opus 5 launched the same month Fable 5 returned to full public availability worldwide on July 1, after a 19-day pause triggered by US export controls and jailbreak concerns. The restore added cybersecurity classifiers as the strongest safeguards yet, without ID-verification requirements. In that context, Opus 5 is not replacing Fable 5, it is the cost-rational alternative for teams that want most of the peak quality without the peak price, while Fable 5 stays the ceiling.

Where it sits in the pricing war

Anthropic's move lands directly against OpenAI's GPT-5.6 launch the same week, which cut Luna prices by 80% and Terra by 20%. Opus 5's $5/$25 sits in the middle: more expensive than the discount tiers, cheaper per task than Fable 5 and GPT-5.6 Sol for agentic coding work. For Claude API users the practical question is simple: run hard tasks on Opus 5 with high effort, reserve Fable 5 for the hardest long-horizon work, and use Sonnet 5's introductory $2/$10 pricing, available through August 31, for high-volume routine calls. A fast Opus 5 endpoint also appeared in provider listings the same day, which gives latency-sensitive teams a second serving option at the same model quality.

The week also showed Anthropic investing in its own compute future: a capacity agreement with AMD for up to 2 gigawatts of MI450 and Helios-generation compute, including up to $5 billion in AMD equity. The deal matters for model buyers because it reduces Anthropic's dependence on a single GPU vendor, which is one of the reasons the company has been able to hold Opus pricing flat while competitors cut and raise tiers around it.

The AMD compute deal and what it signals

The week of the Opus 5 launch, Anthropic and AMD struck a capacity agreement giving Anthropic access to up to 2 gigawatts of MI450 and Helios-generation compute, a deal that includes up to $5 billion in AMD equity. The same week also saw reports of separate compute-lease talks between Meta and Anthropic valued at roughly $10 billion over two years. The strategic picture is clear: Anthropic is diversifying away from single-vendor GPU dependence at exactly the moment it is shipping four Claude 5-series models in two months. For API customers, the compute deal is a supply-chain signal: when a frontier lab locks in multi-gigawatt capacity across two providers, it is planning for sustained inference demand, not just training runs. That is the kind of commitment that makes the stable $5/$25 Opus pricing credible over a multi-year horizon.

The tokenizer caveat

One practical concern from early adopters: the new tokenizer shipped with Sonnet 5 can consume up to 35% more tokens on some inputs, and early feedback on Opus 5 suggests a similar pattern on verbose documents. Wolfram's early WolfBench evaluation of Sonnet 5 found it slightly under Opus 4.6 performance at higher cost on some workloads, and while Opus 5 has not yet accumulated the same breadth of independent reviews, teams migrating production pipelines should benchmark token consumption on their own document types before projecting cost savings from the $5/$25 rate card. The AMD deal and stable pricing are strong signals, but tokenizer-driven cost increases can partially offset them on text-heavy tasks.

What it means for your AI stack

If you build on the Claude API or use Claude Code, Opus 5 changes the default economics. The effort toggle effectively gives you a multi-speed model on one price point, and the benchmark deltas justify re-testing workflows that previously fell back to Fable 5. The watch item is the tokenizer: Sonnet 5's new tokenizer can consume up to 35% more tokens on some inputs, and if Opus 5 shares that trait, per-token savings can partially evaporate on verbose documents. Test with your own prompts before migrating large pipelines. Teams should also benchmark whether the effort toggle at high settings consumes more output tokens than the per-task cost savings imply, since the dynamic between effort level and token burn is still being characterized by early adopters.