Claude Opus 5.5 at $4 per Million Tokens, 1M Context

Published September 23, 2026 - San Francisco, California. Anthropic released Claude Opus 5.5 on September 22, 2026 at $4 per million input tokens and $20 per million output tokens - 40% cheaper than Opus 5 (Anthropic, September 2026).

Opus 5.5 ships with adaptive thinking always on and the strongest alignment test result Anthropic has published. The 1M context and 128K standard output limits are unchanged from Opus 5.

At a glance

• Released September 22, 2026; API model id `claude-opus-5-5`; default Opus in Claude Code 2.1.280.
• Standard direct API rates: $4 input / $20 output per million tokens; cache reads $0.20 per million.
• 40% cheaper than Opus 5 ($5/$25) at input and output, with lower compute to serve.
• Performs at the level of Claude Fable 5.1 for most tasks; leads Anthropic's agentic coding, computer use, and knowledge work evaluations.
• Adaptive thinking always on, default effort `medium` (Opus 5 defaulted to `high`). 1M context / 128K standard output, unchanged.
• Strongest alignment-test score of any Anthropic model; first model since the September 12, 2026 "pacing the frontier" call.
• Sonnet 5.5 and Haiku 5.5 announced for the coming weeks.

Data last verified September 23, 2026 from Anthropic, the Claude pricing page, the Claude Code release notes, and METR/Frontier Design external evaluator summaries.

Quick Answer

Claude Opus 5.5 is Anthropic's new flagship, released September 22, 2026 at $4 per million input tokens and $20 per million output tokens - 40% cheaper than Opus 5. Adaptive thinking is always on. The 1M context and 128K standard output limits are unchanged from Opus 5 (Anthropic, September 2026).

It performs at Claude Fable 5.1 level for most tasks. Anthropic also reported the strongest alignment-test score it has published.

What changed in Opus 5.5 vs Opus 5

Source: Anthropic Claude API pricing page and migration guide (September 2026).

Why the price dropped

Anthropic's release notes say Opus 5.5 requires less compute to serve than Opus 5, and the pricing reflects that. The 40% cut applies at input, output, and cache read tiers.

Anthropic's September 12, 2026 "pacing the frontier" essay set the public framing, and Opus 5.5 is the first model shipped under that posture (Anthropic, September 2026).

For engineering teams, an agentic workload costing $1,000 on Opus 5 today will cost roughly $600 on Opus 5.5 at the same volume. Teams that pinned Opus 5 for cost reasons can revisit the math before migrating.

What the benchmarks show

Source: Anthropic model card for Opus 5.5 (September 22, 2026). Independent benchmarking from Artificial Analysis and METR was in progress at publication.

The 5.5 family roadmap

Anthropic says Sonnet 5.5 and Haiku 5.5 are scheduled for the coming weeks. The pattern follows the August 2025 Fable 5.1 / Mythos 5.1 split.

Sonnet 5.5 will likely target the bulk of agentic coding volume. Haiku 5.5 will likely replace Anthropic's cheap classification tier. Pricing for Sonnet and Haiku 5.5 has not been announced (Anthropic, September 2026).

How to pin Opus 5.5 today

The Claude Code 2.1.280 release notes default Opus to `claude-opus-5-5`. Pin the npm package `@anthropic-ai/claude-code` to 2.1.280 or newer. Direct API callers should set the `model` parameter to `claude-opus-5-5`.

Anthropic recommends starting at the new default `medium` effort. Compare completed tasks, billable tokens, and correction time before replacing Opus 5 in production traffic (Anthropic, September 2026).

Where Opus 5.5 is most cost-effective

The biggest per-token savings land on workloads that re-read the same prompt or system block. Coding agents that keep a 20K-token system prompt in context across a multi-hour session benefit most - the cache read tier is 60% cheaper than Opus 5.

Long-horizon agentic workflows also benefit. Opus 5.5's 20% improvement in task completion at the same token budget means fewer re-rolls for complex coding tasks. Anthropic's published SWE-bench Verified score of 96% puts it ahead of Opus 5 on real software engineering tasks (Anthropic, September 2026).

Document analysis is a third area where the price cut matters. The 1M context window and 128K output limit are unchanged, but a 20-30 page PDF analysis that cost $1.20 on Opus 5 now costs roughly $0.72 on Opus 5.5 at the same volume.

For workloads that are mostly short generation - small completions, single-turn summarization - Opus 5.5 is cheaper but the absolute savings are small. Sonnet 5.5 will likely beat Opus 5.5 on price/performance for those workloads when it ships.

What teams should watch

Anthropic's migration guide notes that some requests that worked on Opus 5 may fail on 5.5. Test agent interfaces before production migration.

Cache reads are 60% cheaper. Long-context agents that re-read the same system prompt will see the largest per-token cost drop.

The knowledge cutoff moved from May to June 2026. Small win for tasks touching late-Q2 facts.

The default thinking effort moved from `high` to `medium`. Opus 5 callers should benchmark at medium before re-enabling high, since the new default is cheaper.

How it compares to the field

OpenAI's GPT-5.5 Pro launched April 23, 2026 at $60 input / $480 output per million tokens - 15x Opus 5.5 on input and 24x on output (OpenAI, April 2026).

Google's Gemini 3.8 Flash and Meta's Muse Spark 1.3 target different tiers - Gemini Flash at consumer speed, Muse Spark on agentic coding. So Opus 5.5 sits at the upper-frontier coding/agentic tier for now (Google; Meta, September 2026).

Sonnet 5.5 will likely compete with GPT-5.6 and Muse Spark directly when it ships.

Migration playbook from Opus 5 to 5.5

Anthropic's migration guide recommends a four-step rollout for teams moving production traffic from Opus 5 to Opus 5.5.

Step one: pin Claude Code to the 2.1.280 release or newer. This guarantees the default Opus model is 5.5 for any new session. Teams that explicitly select the model should also update their `model` parameter to `claude-opus-5-5`.

Step two: run a parallel-evaluation week. Hold 10% of production traffic on Opus 5 and 10% on Opus 5.5 with identical prompts and tools. Compare task completion rate, billable tokens per task, and human-correction rate.

Step three: review the migration guide's known-issues list. Anthropic documented that some Opus 5 success paths now fail on 5.5, and that some agent interfaces lose progress updates. Test these on representative workloads before flipping the default.

Step four: cut over the default and monitor for two weeks. The default thinking effort moved from `high` to `medium`, so teams that relied on `high` should benchmark at `medium` first. Sonnet 5.5 will likely absorb the cost-sensitive portion of Opus workloads once it ships.

Internal links

For the Claude pricing ladder across all Anthropic models, see our Claude API pricing 2026 guide.

For the three-way comparison with GPT-5.5 and Gemini 3.8, see our ChatGPT vs Claude vs Gemini comparison.

For the Fable 5.1 / Mythos 5.1 release context, see our Fable 5.1 anti-distillation and Fable 5.1 cache read price cut posts.

What to expect from the next six weeks

The Opus 5.5 launch is the first of three models in Anthropic's 5.5 family. Sonnet 5.5 is expected in October 2026 and will likely target the bulk of agentic coding volume at a lower price tier.

Haiku 5.5 follows, likely in November 2026. It will replace Haiku 4.5 in classification and routing workloads - and probably drop the per-token price another 50% relative to Sonnet 5.5.

The 5.5 family timing lines up with Anthropic's September 12, 2026 pacing-the-frontier essay. Independent evaluators METR and Frontier Design signed off on Opus 5.5 before release, and the alignment-test score is the strongest Anthropic has published.

For workloads that hit the 1M context ceiling on Opus 5, Opus 5.5 is a direct drop-in with the same ceiling. For workloads that previously pinned Haiku 4.5 for cost reasons, the wait for Haiku 5.5 is the practical move (Anthropic, September 2026).