What Is Claude Opus 5.5 — How Does It Differ from Opus 5 for Agent Coding?
Key takeaway
Claude Opus 5.5 (model id claude-opus-5-5, shipped 2026-09-22) is the upgrade question for people whose coding-agent daily driver is still Opus 5. Officially: 1M context / 128K max output, $4 / $20 per MTok in/out (cache reads $0.20), adaptive thinking always on, default effort medium. Versus Opus 5: about 40% cheaper on typical workloads and 30%+ faster output, with Fable-level performance on most work. This post is not a bench dump—it is when to switch and when not to.
Facts are from Anthropic’s Claude Opus 5.5 page and What’s new in Opus 5.5. “Feel” notes rewrite public reviewer angles; they are not a YouTube transcript.
Why did Opus 5.5 ship?
One-line answer: To bring Fable-class long-running agent/coding quality at a better cost and speed than Opus 5.
Long agent loops force a daily choice between “smart but expensive/slow Opus” and “faster but a tier down.” Opus 5.5 is aimed at that gap. Anthropic’s positioning: near-Fable on most work, with better typical cost and output speed than Opus 5. The useful angle is not the launch headline—it is whether your coding-agent daily driver should move.
How do you summarize the specs?
One-line answer: Lock model id, context, pricing, and thinking/effort in a table. Skip bench tourism here.
| Item | Opus 5.5 (official) | Note vs Opus 5 |
|---|---|---|
| Model ID | claude-opus-5-5 | Prior Opus 5 line |
| Release | 2026-09-22 | — |
| Context / max out | 1M / 128K | — |
| Price (MTok) | $4 in · $20 out | Opus 5: $5 / $25 (20% lower I/O) |
| Cache reads | $0.20 | 60% below Opus 5’s $0.50 |
| Typical workload | ~40% cheaper · ~30%+ faster out | Official guidance |
| Thinking | Adaptive thinking always on | Cannot disable (see breaking) |
| Default effort | medium | Tune in API / products |
| Quality stance | Fable-level on most work | Official positioning |
Tip: Prices and cache figures are Anthropic’s published numbers only. Plan quotas and UI labels differ by product—match Cursor or Claude Code settings on your side.
What breaks when you move Opus 5 → 5.5?
One-line answer: Five highlights: thinking cannot be disabled, forced tool use errors, thinking blocks bound to model+conversation, old computer tool rejected, inter-tool text moves into thinking (quiet streams).
If you run a harness or custom client, swapping the model id is not enough. Fix these from What’s new first:
- Thinking cannot be turned off.
thinking: disabledor enabled+budget_tokens returns 400. Adaptive thinking is always on. - Forced tool use returns an error. Retire “force this tool only” call patterns.
- Thinking blocks are tied to model + conversation (preserved thinking / append-only). Do not replay them across models or reshuffle freely.
computer_20251124is rejected on the Claude API and GCP. Use a newer computer toolset such ascomputer_toolset_20260801.- Text between tool calls lands in thinking blocks. With default display omitted, the stream can look quiet—check summarized display options.
Warning: Even when Cursor or Claude Code picks schemas for you, proxies, MCP bridges, or pinned old computer tools still hit these breaks. Changing only the dropdown and skipping logs often looks like “the model is stuck.”
What strengths show up in agent practice?
One-line answer: Long loops, frontend/UI polish, and workflows that prefer a repo contract (SSOT) over a mega-prompt—these make it a daily-driver candidate.
Specs alone do not tell you when fewer babysitting turns happen. Rewriting public reviewer/field angles, three themes recur:
- Long agent loops: explore → edit → verify across many turns with fewer broken threads or repeated mistakes. That matches the “Opus-grade writing/reasoning plus Fable-grade coding in one daily driver” expectation (e.g. Jarred Sumner-style commentary).
- Frontend / UI polish: layout, component cleanup, and “it works but looks rough” finishing passes where reviewers report less hand-holding.
- Repo contract / SSOT first: put rules, schemas, tests, and agent docs in the repository as the contract, instead of stuffing every norm into a system prompt. Fits workflows where the repo is the source of truth.
If you only ask one-shot chat questions, the delta can feel small. Upgrade value scales with how much work you actually delegate to an agent.
When should you not move yet?
One-line answer: Stay on Sonnet/Haiku for cost/speed-only work; skip Opus 5.5 for short consumer UI/short-form, and for chat where a lecturing tone gets in the way.
| Signal to stay | Why |
|---|---|
| Bulk short edits / simple refactors | Opus-class is overkill; Sonnet/Haiku win on cost |
| Consumer UI / short-form copy | Often cited as a weak spot; tone and cadence may not fit |
| Brainstorm-only chat | If a “lecturing / conservative” tone annoys you, keep a lighter chat model |
| No time to fix API breaks | Leftover thinking-disable / forced-tool / old computer patterns will page you |
| Stable Opus 5 harness already | If the felt gain is tiny, id-only swaps buy risk for little reward |
Move first if you daily-drive long coding-agent loops; otherwise split models by workload.
How do you set model id and effort in Cursor, Claude Code, and the API?
One-line answer: Use claude-opus-5-5 everywhere; leave thinking on; start at effort medium and raise only for hard agent/design work.
- API: set
modeltoclaude-opus-5-5. Do not sendthinking: disabled. Keep default effortmediumunless the task truly needs more. - Claude Code: pick Opus 5.5 in the product UI (labels vary by build) and confirm the wire id is
claude-opus-5-5once in logs. - Cursor: select Opus 5.5 in the agent/model picker. With custom routers (e.g. OpenRouter), verify provider id mapping; on failure, inspect router schema and computer tool versions before blaming the model.
- Homegrown agents: remove forced tool use, move to
computer_toolset_20260801+, keep thinking blocks append-only per conversation, then switch.
{
"model": "claude-opus-5-5",
"effort": "medium"
}Tip: A quiet stream often means inter-tool text is inside thinking with display omitted—not a hung model. Check What’s new on display behavior first.
One-line wrap (next: GPT-6 Astra)
One-line answer: Near-Fable quality with better cost/speed than Opus 5—strong coding-agent daily-driver candidate. Keep Sonnet/Haiku, short-form, and tone-sensitive chat elsewhere.
Do not swap the id alone: clear thinking/computer/forced-tool breaks first. Official sources: anthropic.com/claude-opus-5-5 and What’s new in Opus 5.5.
Next, same “when do you attach it to an agent?” angle for GPT-6 Astra (computer use, long loops) versus everyday sibling Sol. Opus 5.5 leans repo/coding loops; Astra is closer to handing over OS and browser.
If you are also comparing Cursor or Claude Code subscription paths, the step-by-step Gamsgo discount & usage hub may help—read it as a third-party marketplace, not an official reseller.