Repository navigation
Add claude-haiku-5-5 to the model catalog (adaptive thinking, image input, 1M context, tiered pricing) #2892
Description
Activity
Status of the #2892 checklist (delivered in #2911, stacked on #2913):
- Catalog:
claude-haiku-5-5on Anthropic, Amazon Bedrock (on-demand + global/us/eu/jp/au), OpenCode, OpenCode Go, OpenRouter (+ batch), Vercel AI Gateway and Venice. Every row has adaptive thinking, text+image input, 1M context, 128K output, and the request-wide >100K tier at 5x its own base. - Thinking request: adaptive, no
budget_tokens, effort low..max; per-message effort reaches the wire once fix(ai): let the per-message Anthropic effort reach the wire #2913 lands (Anthropic per-message effort never reaches the wire: demoteUnavailableToolReferences drops the content-less effort markers #2912).offstays hidden: whetherdisabledis rejected is unverified and is tracked in Verify Claude Haiku 5.5 tool_reference / tool search against the live API, then enable #2914. - Haiku substring checks (
defaultSupportsToolReferences, native tool search): kept off for Haiku 5.5. Anthropic's table lists it, but it lists Haiku 4.5 identically, and 4.5 rejectstool_reference(#6474). A non-matching rejection would fail the turn. No live credential was available. Tracked in Verify Claude Haiku 5.5 tool_reference / tool search against the live API, then enable #2914 (one live probe each, then enable; the version-aware check is written down there). - Aliases: senpi has no alias table. A bare
--model haikuresolves to the highest undated id, which isclaude-haiku-5-5, proven on the wire in QA. - Recommended / utility models: no recommended ladder names a Haiku. The websearch hosted-search default stays
claude-haiku-4-5(5.5 thinks by default and the search request has no thinking control). The latency/cost comparison is tracked in Verify Claude Haiku 5.5 tool_reference / tool search against the live API, then enable #2914. - Desktop picker: the desktop lists models through the engine's
get_available_modelsRPC (OmoModelDiscovery.ts), so Haiku 5.5 appears with the next senpi runtime pin. No desktop code change is needed.
#2911 uses
Refs #2892. This issue is closed by hand after #2913 and #2911 merge, with the open items carried by #2914.- Catalog:
Delivered by #2911, merged as c09e009 and stacked on #2913.
claude-haiku-5-5is in senpi main and ships with the next release.What ships
- 13 catalog rows across 7 providers: Anthropic, 6 Bedrock routes, OpenCode, OpenCode Go, 2 OpenRouter rows, Venice and Vercel.
- Thinking and pricing: adaptive thinking with effort, image input, and the tiered price bands (the long-context rate above 100K input).
- Default window: 100K, where the 5x long-context price band starts, so nothing past 100K is billed at 5x without an explicit choice.
- Output defaults to 32K so compaction keeps its order: adaptive compaction at 60K, emergency prune at 64.6K, hard limit at 84K.
- To use 1M, set
contextWindow: 1000000, maxTokens: 128000forclaude-haiku-5-5undermodelOverridesinmodels.json. A first-class setting is tracked in Per-model 1M context opt-in for Claude Haiku 5.5 (and other tiered models) #2916.
- API rules covered by request-shape tests: no temperature, top_p, top_k,
budget_tokensor assistant prefill on the Anthropic and Bedrock rows. Thinking is billed as output. The top-level effort stays stable across turns, so the prompt cache holds. Response blocks are parsed by type.
Follow-ups
- Verify Claude Haiku 5.5 tool_reference / tool search against the live API, then enable #2914: a live probe of tool search and tool references, which stay off until then.
- Per-model 1M context opt-in for Claude Haiku 5.5 (and other tiered models) #2916: the 1M opt-in setting, and decoupling the compaction output reserve from
maxTokens. - Expose real thinking-off for Claude Haiku 5.5 (disabled thinking at <= high, no differing per-message effort) #2927: a real thinking-off mode.
- openai-completions forwards temperature/top_p/top_k to Claude models that reject sampling params (OpenRouter :batch, Venice, OpenGateway) #2928: sampling parameters on the openai-completions Claude rows.
Evidence at the merged head d688c38
- Build,
bun run checkand the fullpackages/aisuite: 372 files, 3709 passed, 0 failed. - GitHub CI: 33 checks green.
- QA mock loop: 60/60.
- Review: two independent passes approved it.
Thanks for the request!
Summary
Add Claude Haiku 5.5 (
claude-haiku-5-5) to the model catalog. It was released on 2026-10-07; upstream oh-my-pi (7af89a4) and pingdotgg/t3code (#16903) added it the same day. Our newest Haiku inpackages/ai/src/providers/dataisclaude-haiku-4-5.Model facts (from the upstream commits and Anthropic's model docs)
budget_tokensis rejected with a 400. Effort ranges from low to max (upstream t3code adds ultrathink); the API default is medium./v1/modelsreports neither reasoning, vision, pricing nor limits for this model. A discovered row would therefore come out as text-only and unpriced, so the catalog needs an explicit patch, not a passive regen.How to add it
Mirror how Opus 5.5 (
1ca1a14) and Sonnet 5.5 (fd4353b) were added:packages/ai/scripts/generate-models.ts;providers/data/*.jsonfor every provider that serves it;.manifest.jsonbump;changes.mdentry;test/anthropic-opus-5-5.test.ts, covering thinking mode, input modalities, limits, and long-context pricing.Checklist
claude-haiku-5-5is in the catalog for every provider that serves it, with adaptive thinking, text+image input, 1M context, 128K output, and both pricing tiers.The thinking request uses adaptive mode (no
budget_tokens), and the effort levels low..max are offered, with medium as the default.Haiku substring checks. Two places treat every model whose id contains
haikuas an older Haiku, and need checking against Haiku 5.5's actual capabilities before the model is offered:packages/ai/src/utils/prompt-cache-ttl.ts:68(defaultSupportsToolReferences: Haiku rejects client-sidetool_referenceblocks);packages/coding-agent/src/core/extensions/builtin/tool-search/native-support.ts:26(native tool search).Replace the substring with a version-aware check if 5.5 supports either feature.
Aliases. senpi currently has no bare
haiku/opus/sonnetalias table, unlike t3code, which points its bare aliases at the 5.5 line. If one is added (or if the desktop or omo maps such names),haikushould resolve toclaude-haiku-5-5.Recommended models. Decide whether Haiku 5.5 belongs on any recommended ladder or title/utility-model default that currently names Haiku 4.5.
Desktop model picker (code-yeongyu/omo-desktop-app) follows the catalog: Haiku 5.5 is listed with its effort levels and context options, and Haiku 4.5 is shown as the older model. This goes in the same change set or right after the next runtime pin; no separate desktop issue.
Related
7af89a4, feat(models): add Claude Haiku 5.5 and retire Sonnet 5 and Opus 5 to legacy pingdotgg/t3code#169031ca1a14(Opus 5.5),fd4353b(Sonnet 5.5)