The launch number is a vendor claim.
Anthropic says Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. It lists cache reads at $0.20 per million tokens. The company also reports a one-million-token context window, 128,000-token maximum output, and medium default effort.[1][2]
The launch post reports gains on agentic coding and other benchmarks. It also says benchmark margins have become a less reliable guide to real work. That caveat belongs beside the scores, not under the carpet.[1]
A leaderboard can nominate a model. Only a matched run can approve your swap.
The harness has its own release.
Claude Code 2.1.280 made Opus 5.5 the default Opus model. The same release added a control for the maximum MCP description length and changed telemetry around oversized hook output. It also fixed policy checks for writes through symlinked paths and rebuilt damaged cached message lists in one fullscreen failure case.[3]
OpenCode's GPT-6 support tells the same story from another shop. Release 1.18.30 added an Astra-specific system prompt, updated provider SDKs, and added reasoning-effort variants for supported GitLab models.[4]
Those are not cosmetic adapters. A system prompt changes instructions. An SDK changes request and response handling. Effort changes how much work the model spends. Tool descriptions change what the model can see. A model ID without those facts does not identify the system you tested.
Context size is capacity, not permission.
Anthropic's model documentation lists Opus 5.5 with adaptive thinking, medium default effort, a one-million-token context window, and tool use. The same documentation separates API clients from agent runtimes. Claude Code and the Agent SDK add the loop and tool execution above the raw model API.[2]
A larger window can hold more repository state. It can also ingest more stale files, generated output, secrets, or irrelevant history. Keep retrieval rules and tool permissions under review. Do not turn a larger fuel tank into a wider loading dock.
Run a boring swap test.
- Pin the exact model identifier and provider route. Record whether an alias can move.
- Pin the harness version. Save the system prompt or vendor release note that describes its model-specific path.
- Record effort, context cap, tool list, MCP description cap, retry limit, and spend ceiling.
- Replay a small set of tasks with known patches and known failure modes.
- Compare patch correctness, review time, tool calls, wall time, input, output, and cache cost.
- Promote the new setup only if the receipt beats the old one on the measures your work needs.
The exciting model should earn its key on boring evidence.