nntrivi2001
9c573f1e3c
Decisions, environment quirks and pitfalls learned while working on AgentHub (docs/claude-memory/, indexed by MEMORY.md). Notes holding personal account details are left out. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2.0 KiB
2.0 KiB
name, description, metadata
| name | description | metadata | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| claude-haiku-5-5 | Claude Haiku 5.5 (2026-10-07) facts + its 100k-token price tier, which AgentHub's pricing does not model |
|
Source: https://www.anthropic.com/claude-haiku-5-5 (read 2026-10-08).
- Model id
claude-haiku-5-5, released 2026-10-07. Cheapest/fastest small model; first Haiku with effort levels (low/med/high/xhigh/max); computer/browser use in beta; new tokenizer uses slightly more tokens per task. - Per MTok, split at a 100k-token prompt: input $0.10/$0.50, output $0.50/$2.50, cache write $0.125/$0.625, cache read $0.01/$0.05. (Haiku 4.5: $1/$5/$1.25/$0.10, flat.)
- Same day: Sonnet 5.5 cache read was halved, $0.20 → $0.10 (input $2, output $10 unchanged).
- Anthropic positions it for summaries, compaction, classification, subagents; Sonnet/Opus 5.5 still better for complex agentic coding (Terminal-Bench 4.0: 39.2% vs Sonnet 70.6%).
- Platforms: Claude API, AWS, Google Cloud, Azure.
AgentHub status 2026-10-08: the >100k tier is now modelled (uncommitted), as SUBSET columns long_*_tokens on token_usage/token_usage_hourly plus model_pricing.long_input_per_mtok/long_output_per_mtok, synced from LiteLLM's *_above_100k_tokens. It was deliberately NOT done as a new primary-key dimension: prod (4820) and dev (4830) share one DB, so a PK rebuild would break the older build's ON CONFLICT, and splitting rows would double-count under the high-water-mark baselines. Threshold is the single LONG_CONTEXT_TOKENS = 100k.
Useful for the user's own work:
- A good cheap model for subagents doing exploration, summaries or classification. AgentHub's AI commit message (
server/lib/commit-message.js,--model haikualias) now gets it automatically and costs about 10× less. Nothing needs to change. - Its cyber safeguards still BLOCK pentest/attacker techniques, so it is the wrong model for Strix pentest work. Use Sonnet/Opus there.