Release facts and lifecycle
- Release announced
- October 7, 2026
- Model ID
- claude-haiku-5-5
- Announced context
- 1 million tokens
- Announced maximum output
- 128,000 tokens
- Thinking
- Adaptive thinking with effort control
Original provider sources checked . These are source facts, not a benchmark of Flonno's deployment.
Open the live model record, benchmarks and release journal →Where to start
Test a representative queue of easy and difficult requests. Tune the native effort parameter and record tail latency, schema validity, incorrect values and retry cost.
- High-volume workflows
- Structured extraction
- Latency-sensitive tasks
What to check before deployment
Migration from Haiku 4.5 needs attention: manual budget_tokens thinking, sampling parameters and assistant prefill can return 400 errors on Haiku 5.5.
Confirm the exact model ID, checkpoint revision and endpoint in your test record. A base model name does not establish a provider's quantization, accepted inputs, context limit or current availability.
Measure useful results
Build a small evaluation from the workflow you already run. Define the expected answer before testing. Track correct results, formatting errors, retries, complete request time and reported input and output usage.
Count rejected answers and manual corrections. A low token rate can still produce a high cost per accepted result when the workflow needs retries or repair.
Estimate your token budget →Sources and scope
This page is an editorial evaluation guide, not a benchmark result or a claim that this model is served by Flonno.
Official model announcement ↗License / access: Provider service terms. Read the linked license or service terms for the full conditions. Original release and lifecycle sources checked 2026-10-10.