Amp moves smart mode to Claude Opus 4.5
Amp has switched the default model in its smart mode to Claude Opus 4.5, two days after rolling out Gemini 3. The change follows a brief run with Gemini 3 that delivered strong peak performance but also produced frustrating edge cases and costly behavior for some workflows. Opus 4.5 aims to match Gemini 3’s capabilities while reducing those low-end failures, improving consistency, and trimming latency.
Performance and cost snapshot
Amp’s internal comparisons place Opus 4.5 ahead on several fronts:
| Sonnet 4.5 | Gemini 3 Pro | Opus 4.5 | |
|---|---|---|---|
| Internal Evals | 37.1% | 53.7% | 57.3% |
| Avg. Thread Cost | $2.75 | $2.04 | $2.05 |
| 0–200k Tokens Only | $1.48 | $1.19 | $2.05 |
| Off-the-Rails Cost | 8.4% | 17.8% | 2.4% |
| Speed (p50, preliminary) | 2.4 min | 4.3 min | 3.5 min |
Key points from the numbers:
- Off-the-rails cost is materially lower for Opus (2.4%), indicating fewer expensive failure modes.
- Median responsiveness (p50) for Opus is faster than Gemini 3 Pro in preliminary measures.
- Average thread cost for Opus is similar to Gemini 3 Pro overall, but cost behavior varies by context length.
Context limits and guidance
Opus is currently limited to 200k tokens of context. For workflows that require longer context, Sonnet can still be used by enabling "amp.model.sonnet": true or the --use-sonnet CLI flag. Amp emphasizes small threads as the recommended pattern for both quality and cost; guidance on managing context is available in the context-management guide.
When Opus is the better fit
- Long threads (200k+ tokens): Opus will be much cheaper if long-context support is not required; if Sonnet’s long context is necessary, the Sonnet setting remains available.
- Capability ceilings: When Sonnet or Gemini encounter limits, Opus generally provides higher accuracy and can reduce token wastage, lowering overall costs.
- Comparing to Gemini 3 Pro: Opus tends to be ~40% more expensive but offers faster responses and greater tolerance for ambiguous prompts.
- Comparing to Sonnet 4.5: Opus is ~35% pricier for straightforward tasks, but it performs better on more difficult prompts where Sonnet may struggle.
Amp’s decision reflects a trade-off between peak brilliance and real-world stability: Opus 4.5 aims for steadier, more cost-efficient behavior, with fewer token-wasting mistakes and higher cache hit rates that reduce latency and human intervention.
Read the original announcement: https://ampcode.com/news/opus-4.5