Technical information was last verified on April 17, 2026. The AI/LLM field moves fast — re-check official docs if more than 6 months have passed.
Who should read this
Summary: Claude Opus 4.7 officially launched on April 16, 2026. This article analyzes six key changes from a developer’s perspective: Task Budgets, a new tokenizer, the Adaptive Thinking transition, /ultrareview multi-agent code review, sampling parameter breaking changes, and agent performance improvements.
This article is for developers using the Claude API in production, teams designing agent workflows, and engineers planning migration from Opus 4.6 to 4.7. This covers official changes from the production release, not leak analysis.
The 6 key changes in detail
1. Task Budgets — token budget control for agent loops
The most noteworthy new feature in Opus 4.7. Task Budgets let you specify an approximate upper bound on the tokens an agent can consume to complete a single task.
Core mechanics:
- Pre-set budget: Specify a token target for the entire agent loop at API call time
- Real-time countdown: The model monitors its remaining budget in real time as it works
- Autonomous adjustment: As the budget shrinks, the model narrows its exploration scope and converges toward a conclusion
Previously, controlling token consumption in agent loops required hard cuts from the outside. Because the model itself is budget-aware with Task Budgets, it can wrap up work naturally while still keeping costs under control.
2. New tokenizer — cost estimation re-validation required
Opus 4.7 uses a new tokenizer. Processing the same text now consumes 1x to 1.35x the tokens compared to before. That means up to 35% more tokens may be needed.
Impact areas:
- Cost estimation: Existing token-count-based cost calculations become inaccurate
- Context window management: The same text occupies more tokens, changing context utilization efficiency
- Rate limit calculations: Throughput may decrease if you use per-minute token limits
3. Adaptive Thinking — Extended Thinking fully replaced
This is the highest-impact change in this release.
- Thinking responses are omitted by default: Calling without special configuration means thinking content is not included in the response
- Extended Thinking budget has been removed: The previous thinking budget parameter no longer works
- Adaptive Thinking is the sole thinking mode: Instead of enabling Extended Thinking, you activate Adaptive Thinking and the model automatically adjusts thinking depth based on problem complexity
According to Anthropic, Adaptive Thinking reliably outperforms Extended Thinking. Their assessment is that letting the model adjust thinking depth as needed is more efficient than allocating a fixed thinking budget.
All code that explicitly set Extended Thinking budgets needs to be updated. Either switch to Adaptive Thinking or explicitly disable thinking.
4. Breaking change — sampling parameter restrictions
Setting non-default values for temperature, top_p, or top_k now returns a 400 error.
This is not a warning — it is an immediate error. Every API call that used these parameters will fail the moment you switch to Opus 4.7.
The fix is straightforward:
- Remove temperature, top_p, and top_k parameters from your API calls entirely
- Avoid explicitly setting defaults (e.g., temperature=1.0) — omitting the parameter altogether is safest
5. Agent performance — qualitative gains in multi-step and long-horizon reasoning
Opus 4.7 is an incremental upgrade over Opus 4.6 (1M token context window), but delivers meaningful improvements in agent execution stability.
- Stronger multi-step task performance: Failure rates in intermediate steps of multi-stage tasks have decreased
- More stable agent execution: Tool call ordering and result interpretation in tool-dependent workflows are more accurate
- Improved long-horizon reasoning: Logical consistency across extended contexts has improved
This is not about benchmark numbers but about practical improvement — the cases where agents “occasionally behaved strangely” happen less often. The more complex your tool chains, the more you will feel the difference.
6. /ultrareview — multi-agent code review
Alongside Opus 4.7, Claude Code gains the /ultrareview command. This takes a fundamentally different approach from the existing /review.
Existing /review vs /ultrareview:
/review— a single agent reads the diff once and leaves comments. Takes 3-4 minutes/ultrareview— multiple reviewer agents analyze the diff and surrounding code in parallel on Anthropic’s cloud infrastructure. Takes roughly 17 minutes for an 11,000-line PR
How it works:
- Parallel analysis — multiple agents analyze the diff and surrounding code simultaneously. Each agent focuses on a different class of issue (security, logic errors, edge cases, design flaws)
- Independent verification — discovered issues are verified against actual code behavior. Style-level suggestions like “you should use const here” are filtered out; only real bugs and design flaws remain
- Deduplication + severity ranking — when multiple agents find the same issue, duplicates are merged and results are sorted by severity
- Inline comments — results appear as inline comments on the exact code lines where issues exist
Key differentiator: The goal is not simple linting or style checks but detecting the subtle design flaws and logic gaps that a senior reviewer would catch. In Anthropic’s words, it “flags what a careful reviewer would catch.”
Usage limits:
- 3 free trials for Pro and Max plan users (one-time, not monthly renewal)
- Paid usage after that (see Anthropic documentation for specific pricing)
Migration guide: Opus 4.6 to 4.7
Checklist to verify before switching:
Immediate fixes (errors if not addressed):
- Remove temperature, top_p, and top_k parameters from API calls
- Remove Extended Thinking budget configuration or replace it with Adaptive Thinking
Cost impact validation: 3. Measure the token increase from the new tokenizer using your actual workloads 4. Evaluate whether to adopt Task Budgets and set cost ceilings for agent loops
Performance validation: 5. Run eval sets against your critical prompts and check for output quality changes 6. Perform end-to-end testing on your agent workflows
Opus 4.6 vs 4.7 comparison
| Opus 4.6 | Opus 4.7 | |
|---|---|---|
| Token budget control | External hard cuts only | Task Budgets -- model is budget-aware while working |
| Tokenizer | Previous tokenizer | New tokenizer (1x-1.35x token increase) |
| Thinking mode | Extended Thinking (manual budget setting) | Adaptive Thinking (auto-adjusting, thinking omitted by default) |
| Sampling parameters | temperature, top_p, top_k available | Non-default values return 400 error |
| Multi-step performance | Baseline | Meaningful stability improvement |
| Long-horizon reasoning | Baseline | Improved long-range consistency |
| Context window | 1M tokens | 1M tokens (unchanged) |
| Claude Code review | /review (single agent) | /review + /ultrareview (multi-agent parallel verification) |
| Co-launch | -- | AI design tool (website/presentation generation) |
Mistakes to avoid
Further reading
- Claude Code Desktop App Redesign: When to Switch from CLI and When Not To — Tooling changes launching alongside Opus 4.7
- Harness Engineering: From Prompts to Runtime Control — Agent design principles that remain valid across model changes
- Prompt Version Management for Production AI Services — How to prevent prompt regressions during model transitions