Claude Opus 4.6
Released: February 5, 2026
Model ID: claude-opus-4-6
Pricing: $5 / $25 per million input / output tokens
Status: legacy. Superseded by claude-opus-4-8.md and claude-opus-5-5.md.
Last verified: legacy, written at the Opus 4.6 launch (February 2026)
Key features
- 1M token context window: Process entire codebases, long documents, or
multi-file analysis in a single session. 76–78.3% retrieval accuracy on
MRCR v2 benchmarks.
- Adaptive thinking: Replaced manual extended thinking with dynamic
reasoning depth calibration. Four effort levels:
- Low — speed-optimized, simple tasks
- Medium — balanced for moderate complexity
- High (default) — deep reasoning for production workloads
- Max — spare-no-expense for hardest problems (complex debugging, novel
research)
- Agent teams (research preview): Claude Code orchestrates multiple
parallel agents. A lead agent decomposes tasks, assigns to specialist
teammate agents, synthesizes results.
- Context compaction: Automated API summarizes older conversation parts
when nearing context limits. Prevents degradation in extended sessions.
- 128k output tokens: Generate entire documents or large code
implementations in a single pass.
- Computer use improvements: Higher accuracy navigating software
interfaces — clicking, typing, managing browser tabs.
Migration notes
From Opus 4.5 or earlier
| Area |
Change required |
| Extended thinking |
Replace budget_tokens config with adaptive thinking ({type: "adaptive"}). Set effort parameter. |
| Context window |
1M tokens available — review whether you need context management strategies or can simplify. |
| Output length |
128k output tokens now supported — remove chunking workarounds if applicable. |
| Agent workflows |
Agent teams available — consider decomposing complex migrations into parallel subtasks. |
| Computer use |
If using computer use, retest — behavior and accuracy changed significantly. |
From Opus 4 / 4.1
All of the above, plus:
- Re-check prompt templates — Opus 4.6 follows instructions more literally.
- Verify tool schemas — stricter validation on tool definitions.
- Update model ID from
claude-opus-4-* to claude-opus-4-6.
Quirks and best practices
- Effort level tuning: Start with
high (default). Drop to medium for
high-throughput, low-complexity workloads. Use max sparingly — it consumes
significantly more tokens.
- Context compaction awareness: For long-running agentic sessions, the
model auto-compacts older context. If your workflow depends on exact recall
of early messages, structure important information as system-level
instructions rather than early user turns.
- Agent team decomposition: When using agent teams, keep subtask
descriptions self-contained. Each teammate agent has its own context — don’t
assume shared state between teammates.
- Literal instruction following: More literal than Opus 4.5. Prompts that
relied on the model “reading between the lines” may need explicit
reformulation.
- Long-context retrieval: Despite 1M context, retrieval accuracy can vary
by position. Place critical information at the start (system prompt) or end
(recent messages) of the context — middle positions have lower recall.
- Output planning: With 128k output capacity, the model can over-generate.
Set explicit length or format constraints in your prompts when concise output
is needed.
Model ID reference
| Version |
Model ID |
Status |
| Opus 4.8 |
claude-opus-4-8 |
Current (GA) |
| Opus 4.7 |
claude-opus-4-7 |
Active |
| Opus 4.6 |
claude-opus-4-6 |
Active |
| Opus 4.5 |
claude-opus-4-5 |
Active |