Repository navigation
feat: add google/gemini-4-carbon (tool-verified) - #13
Merged
Merged
Conversation
Gemini 4 Carbon is Google's most capable Gemini 4 reasoning model (complex software engineering, deep math and science, long agent work); Argon is the efficient Gemini 4. Same shape as Argon: 1M context, 65,536 max output, vision. The gateway listing is BlockRunAI/blockrun#1035; until that deploys, blockrun.ai's /api/v1/models does not carry it. Unlike Argon, Carbon has tool evidence: on the BlockRun base gateway a two-round tool history through both the stream and non-stream paths returned Carbon's own next tool call (finish_reason tool_calls), no fallback. It is recorded in probe-tools.json with its source (not probe-model-profiles.mjs, so no latency profile), written supportsTools: true, and joins VERIFIED_TOOL_FALLBACKS, which only offers a model the host's live catalog prices, so nothing routes to it before the gateway lists it. model-capabilities.ts is the sync script's output for the 2026-11-08 catalog plus a Carbon row (verified by running it); a newer live listing also carries anthropic/claude-fable-5.2, left for its own sync. Argon and Mistral Large 4 stay fail-closed, now pinned by a test. Decision snapshot: 0 of 88 primaries, tiers or costs move; the 20 tool decisions gain Carbon as a trailing fallback candidate. README counts and dist/ updated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds
google/gemini-4-carbonto the capability snapshot: 1M context, 65,536 max output, vision, supportsTools: true. Carbon is Google's most capable Gemini 4 reasoning model (complex SWE, deep math/science, long agent work); Argon is the efficient Gemini 4.Tool evidence. On the BlockRun base gateway, a two-round tool history through both the stream and non-stream paths returned Carbon's own next tool call (
finish_reason: tool_calls), with no fallback. That is recorded inscripts/probe-tools.jsonalong with where it came from. It was not run throughprobe-model-profiles.mjs, so there is no latency profile yet, and a re-probe overwrites the entry. Carbon joinsVERIFIED_TOOL_FALLBACKS, the same admission rule as GPT-6.1 Sol and DeepSeek V4.1 Flash (#10).Not live yet. The gateway listing is BlockRunAI/blockrun#1035. Until it deploys, blockrun.ai does not list Carbon.
VERIFIED_TOOL_FALLBACKSonly offers models that are priced in the host's live catalog (modelPricing), so nothing routes to Carbon before then.Generated file.
model-capabilities.tsis exactly whatsync-model-capabilities.mjsoutputs for the 2026-11-08 catalog plus a Carbon row; I ran the script to confirm. The current live listing also hasanthropic/claude-fable-5.2, which I left for its own sync. Argon and Mistral Large 4 are unchanged (fail closed), and a new test now pins both.Decision snapshot. Across all 88 decisions, no primary, tier or cost changes. The 20 tool decisions each gain Carbon as a trailing fallback candidate (
candidates=Ngoes up by 1).Tests:
npm run checkpasses (typecheck, 192/192 vitest, build).dist/is rebuilt.Pricing for hosts: $2/$10 per 1M (cache $0.10) through 2026-11-30, then $4/$20 (cache $0.20), the same as Argon. Router-core takes pricing from the host catalog.
🤖 Generated with Claude Code