Pimp My IDE / garage dispatchBack to the garage
October 7, 2026 / model routing / workload proof

Route the small jobs. Keep the escape lane.

Claude Haiku 5.5 makes a strong case for cheap, fast agent work. A benchmark row does not tell you which production jobs may move.

The right unit of migration is a workload, not a model name.

The price drop is real.

Anthropic released Claude Haiku 5.5 on October 7. The company positions it for classification, extraction, routing, summaries, compaction, browser use, and narrow subagent work. Anthropic says it is its fastest model at standard speed.[1]

The listed API price starts at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Requests above that threshold use higher rates. Anthropic says the average task costs about 75 percent less than Haiku 4.5 after accounting for the new tokenizer and the mix of short and long requests.[1]

The current model reference lists the API ID as claude-haiku-5-5. It reports a one million token context window, a 128,000 token maximum output, adaptive thinking, and a default effort of medium.[2]

Cheap output is useful only when the failure bill stays cheap too.

Benchmark names are workload clues.

Anthropic reports 39.2 percent on Terminal-Bench 4.0 and 72.4 percent on the OSWorld 2.1 offline subset. The same announcement reports a much higher Terminal-Bench score for Sonnet 5.5. Anthropic's own guidance keeps complex agentic coding on the larger models and points Haiku at narrower work.[1]

Terminal-Bench joins a task dataset to an execution harness. Each task has an English instruction, a test script, and an oracle solution. That structure matters more than the headline number. A local model trial needs the same kind of executable pass condition.[3]

OSWorld tests agents inside real computer environments. Its repository warns that missing account or proxy setup can lower scores. Verified public results require the maintainers to run the submitted agent or inspect monitoring data and trajectories. Environment setup is part of the measurement.[4]

Move one route at a time.

Start with work that has a narrow input, a machine-checkable result, and a cheap fallback. Classification and extraction often fit. Repository-wide coding, irreversible actions, and ambiguous incident diagnosis need stronger proof before a smaller model owns them.

Build a held-out set from real traffic. Keep the current model as the control. Record task success, false positives, false negatives, tool errors, p50 and p95 latency, input tokens, output tokens, and total cost. Averages hide the requests that hurt users.

Run the candidate in shadow mode first. Compare results without letting it act. Then canary a small share. Send low-confidence, malformed, or policy-sensitive cases back to the current route. Keep that fallback until the small model has passed the workload under production limits.

Use this route card.

SMALL-MODEL ROUTE TEST
WORKLOAD=[ONE NAMED JOB]
CURRENT_MODEL=[PINNED ID]
CANDIDATE_MODEL=claude-haiku-5-5
HELD_OUT_SET=[REVISION + COUNT]
PASS_ORACLE=[EXACT CHECK]
METRICS=[QUALITY + P50/P95 + TOKENS + COST]
FALLBACK=[TRIGGER + DESTINATION]
STATUS=[TEST PLAN / NOT APPROVED]
Interactive makeover / small-model shunt box

Wire the test before moving traffic.

This replaces a model picker and a migration checklist. One workload selector changes the test card. Four native gates drive a connected route. The final state says only that the test structure is ready.

Select workloadFreeze sampleMeasure resultKeep fallback

Workload selector

Choose one job. Then select each section that your test plan will require.

Choose a workload
Required test plan sections

Route readout

The shunt fills as plan sections are selected. Amber means draft work remains. Green means the structure is filled, not that the model is approved.

Current route to candidate trial0 of 4 selected
Control routecurrent classifier
Candidate triallabel holdout only
Plan stateTemplate incomplete
QualityConfusion matrix and abstentions
Speedp50 and p95 end to end
SpendInput, output, cache, retries
Sources read / October 7, 2026

Source log

Open the evidence boundary
  1. Anthropic, "Claude Haiku 5.5", published and read October 7, 2026. This supplies the release scope, pricing, vendor benchmark results, customer quotes, effort settings, intended workload guidance, and tokenizer caveat.
  2. Claude Platform model overview, read October 7, 2026. This supplies the current model ID, context window, maximum output, latency class, effort default, and listed API prices.
  3. Terminal-Bench repository, read October 7, 2026. This supplies the benchmark's task and harness structure, including instructions, test scripts, oracle solutions, and dataset versioning.
  4. OSWorld repository, read October 7, 2026. This supplies the real-computer environment scope, setup caveats, result artifacts, and verified public evaluation process.
  5. Hacker News item 49996437, resolved through the official API and read October 7, 2026. It was the discovery route. Product and benchmark claims above come from the linked first-party sources.

Evidence boundary: vendor benchmark and customer figures were not reproduced. No Claude API key was used. No model request, migration, latency test, or cost comparison ran for this dispatch. The shunt box drafts a workload test plan. It does not score output or approve a production route.