Claude Opus 5.5 Pricing, Benchmarks and When to Switch

Anthropic released Claude Opus 5.5 on September 22, 2026, and the most instructive detail is not the benchmark table — it is the price underneath it. The model arrives at $4 per million input tokens and $20 per million output tokens, twenty percent below the Opus 5 pricing that landed only two months earlier, while an independent measurement from the benchmarking firm Artificial Analysis places it at the top of their intelligence index. Capability rising while unit cost falls is the pattern that decides which models teams can actually afford to leave running in production, and this release is an unusually clean instance of it.

• Pricing: $4 per million input tokens and $20 per million output, down from $5 and $25 on Opus 5; cache reads fall to $0.20 per million from $0.50
• Independent score: Artificial Analysis measured 58 on its Intelligence Index v4.3.2, five points clear of the 53 shared by GPT-6 Astra and Claude Fable 5.1
• Specs: Model ID claude-opus-5-5, a 1M-token context window, 128K max output, and a June 2026 knowledge cutoff
• Availability: Live on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry from day one, with Sonnet 5.5 and Haiku 5.5 slated for the coming weeks

What Does Claude Opus 5.5 Cost?

The full pricing sheet matters more than the headline figure, because agentic workloads are dominated by cached context rather than fresh input. Input runs $4 per million tokens and output $20. Cache writes drop to $5 per million from $6.25, and cache reads — the line item that actually accumulates in a long-running agent loop — fall to $0.20 per million from $0.50, a sixty percent reduction. Anthropic reports that the combined effect is roughly forty percent lower cost on typical workloads, which is a larger drop than the twenty percent headline suggests.

There is a mechanical reason the price could move. Anthropic says the model generates output about thirty percent faster than Opus 5 and requires less compute to serve, so the reduction reflects cheaper inference rather than a promotional rate. That distinction is worth noting for anyone building cost models: efficiency-driven price cuts tend to hold.

How Does Opus 5.5 Compare to Fable 5.1?

This is the genuinely novel part of the release. Fable 5.1, which we covered when it shipped at the start of September, is the larger and more expensive model at $10 and $50 per million tokens. On Anthropic's published evaluations, the smaller Opus 5.5 reports higher numbers across the board — 66.4% versus 55.8% on Terminal-Bench 4.0, 57.8% versus 51.8% on CursorBench 4.0, and 1,846 Elo versus 1,735 on GDPval-AA v2.1.

Those are the vendor's own figures, so the independent corroboration carries the weight here. Artificial Analysis, which runs its own evaluation harness, scored Opus 5.5 at 58 on its Intelligence Index against 53 for Fable 5.1 and 51 for Opus 5, with leading results on six of the index's ten component evaluations including Humanity's Last Exam and SciCode. Two separate measurement methodologies pointing the same direction is a stronger signal than either alone.

Anthropic's own documentation now reflects the shift: it recommends starting with Opus 5.5 for most workloads and reserving Fable 5.1 for demanding reasoning and long-horizon agentic work. A mid-tier model outperforming the flagship above it is a reminder that parameter count and capability have decoupled further than the naming conventions imply.

What Changes for Developers

Beyond price and scores, three practical details stand out. Adaptive thinking is always on and steered through five effort levels from low through max, with medium as the API default — meaning the cost of a request now depends on a parameter as much as on the model choice. The context window holds 1M tokens with 128K maximum output. And Anthropic committed to a retirement date no sooner than September 22, 2027, which gives teams a concrete planning horizon.

Anthropic also focused on prose quality, targeting what it describes as formulaic and convoluted writing by putting important information earlier and reducing jargon. That is harder to benchmark than SWE-bench, but for anyone whose product surfaces model output directly to users, it is arguably the more consequential change.

Why This Release Matters

The broader pattern across our AI coverage is that the interval between frontier releases keeps compressing while cost per unit of capability keeps falling. Opus 5 arrived in late July at half the price of its predecessor; Opus 5.5 repeated the move two months later. Anthropic reports that on its automated behavioral audit of roughly 2,000 scenarios, Opus 5.5 is "the strongest-performing model we've tested to date," and external evaluators METR and Frontier Design reviewed the model before release.

For developers, the practical takeaway is straightforward. The model ID is a drop-in change, the pricing is lower on every line, and two independent measurement approaches agree the capability moved up. That combination does not come along often enough to ignore.

Sources: Anthropic — September 22, 2026; Artificial Analysis — September 22, 2026; TechCrunch — September 22, 2026; The Decoder — September 22, 2026.