fix(amazon-bedrock): correct metadata, pricing, and profiles

[rekram1-node]

Summary

Full audit of all 153 existing Bedrock entries, using five vendor-scoped audits, independent review, AWS/lab documentation, current public AWS price dimensions, and targeted live API checks.

• 48 existing price records corrected, including 23 Claude source-region price quartets, Nova cache-write accounting, and seven Qwen/Gemma/Voxtral price records.
• 126 existing entries have resolved metadata or pricing corrections. Additional changes add citations or replace duplicate inline metadata with base_model.
• 15 missing entries added, bringing Bedrock to 168 entries; eight additions are explicitly marked deprecated/legacy.
• No existing model IDs removed, API routes changed, or lifecycle statuses changed. The previously discussed questionable bare/prefixed IDs remain.

All changes are catalog TOML. New provider files use complete shared lab metadata and host-specific overrides.

Audit coverage

Scope	Existing entries audited
Anthropic	60
Amazon Nova	18
OpenAI	20
Meta / Writer / Mistral	25
DeepSeek / Qwen / xAI / NVIDIA / Google / MiniMax / Z.AI / Moonshot	30
Total	153

Important corrections

Pricing

• Match Claude geography-specific rates to exact AWS source-region Standard dimensions; do not apply a universal 1.1x multiplier. Opus 4.1 stays at its correct $15/$75; AU Opus 4.6's erroneous $16.50/$82.50 becomes $5.50/$27.50.
• Nova cache writes are billed at the input rate, with no additional surcharge. Live usage reports cache-write tokens separately from ordinary input tokens, so omitting their price undercounts processing. Add the full input-equivalent cache-write rate and the missing bare Nova 2 Lite cache-read price.
• Correct Qwen Coder 480B, Coder Next, Next 80B, VL 235B, Gemma 3 12B/27B, and Voxtral Small reference prices.
• Replace inferred Astra pricing citations with its now-published AWS card, which confirms all existing prices and long-context tiers.

Price comments identify the reference source Region. One scalar catalog cost object cannot represent every eligible source Region's rate. Standard/Global/Geo, 1K-versus-1M units, cache TTL, and service-tier dimensions were checked independently.

Metadata and capabilities

• Correct Nova 2 Lite, Palmyra X4, Opus 4.5, Mistral Large 3 and other confirmed release metadata; distinguish lab defaults from Bedrock caps.
• Live boundaries establish Nova Micro/Lite/Pro output 10,000, Nova 2 Lite Bedrock output 65,535 (lab guide remains 65,536), GPT-OSS Runtime output 128,000, and Llama 4 Bedrock output 8,192.
• All six Sonnet 4.6 Bedrock routes accept 128,000 output settings; US and AU reject 128,001 with an explicit 128,000 maximum. Keep this as a host override rather than changing the native lab default.
• Correct vision/attachment/open-weight facts and host-specific modalities, including text-only Llama 3.3 and Writer's native-versus-Bedrock differences.
• Correct missing reasoning controls using actual host fields; verify representative on/off pairs. Gemma 4 uses effort none/high; GPT-OSS/Safeguard and closed OpenAI Runtime models have different wire shapes, documented in the files.
• Live JSON-schema checks confirm US/Global Astra and Luna support; Sonnet 5 rejects constrained format on US/Global Converse. The Astra correction overlaps fix: Remove structured output restriction for Bedrock GPT-6 Astra #6593.
• Reverse documentation-only limit corrections when direct API checks contradicted them; acceptance alone is not presented as proof of an absolute ceiling.

Added entries

Current-generation entries

• in.openai.gpt-5.6-luna, in.openai.gpt-5.6-terra — India profiles; successful live requests from Mumbai.
• us-gov.openai.gpt-oss-20b-1:0, us-gov.openai.gpt-oss-120b-1:0 — exact AWS-published GovCloud profiles and source-region prices; not live-invoked with commercial credentials.
• google.gemma-4-31b, google.gemma-4-26b-a4b, google.gemma-4-e2b — live-verified Mantle Responses routes, reasoning and structured JSON.

Legacy entries, explicitly status = "deprecated"

• us.amazon.nova-premier-v1:0.
• Sonnet 4: us., eu., apac., global. profiles.
• Haiku 3: us., eu., apac. profiles.

Legacy invocations can require use within the previous 30 days. Haiku 3 reaches its documented EOL on 2026-09-10; Sonnet 4 on 2026-10-14. These additions record existing profiles for eligible users and are not represented as new active models.

Deferred / unresolved

• Withdrew tentative Llama 3.2 additions after five profiles returned explicit end-of-life errors. An ACTIVE profile listing alone was insufficient.
• Deferred new Mantle Chat IDs with client API-selection dependencies, Daybreak variants lacking complete control/cap evidence, and unresolved NVIDIA GovCloud reasoning variants.
• Existing NVIDIA Nano3/Super3 controls and some exact Mistral/open-model output ceilings remain unresolved. Existing values are retained where sources conflict; the audit is not a claim that every field was freshly live-verified.
• Existing questionable routes remain for the separately requested removal discussion.

Sources and validation

Each changed model's leading comments cite its primary sources. Principal references:

• AWS model cards
• AWS public Standard pricing
• AWS Foundation Models pricing
• Nova model guide
• Bedrock cache billing
• Astra card

Checks completed:

• bun validate passes on the combined final catalog.
• Generated-catalog comparison against an immutable baseline: 0 removals, 0 existing route changes, 0 existing status changes.
• Verified saved pricing against 4,448 current AWS product price records, with no differences; independently matched 262 Claude dimensions by exact rate code/region/unit.
• Reviewed the 39 resolved records on other providers affected by shared lab corrections (dates, Llama 3.3 attachment, Nova limits and inherited metadata).
• git diff --check passes; changed-path check confirms catalog TOML only.

[github-actions]

Action items

• [high] [violation] models/mistral/magistral-small-2509.toml:17 - Check: New lab metadata must be complete (limit.context + limit.output, plus required capability fields). Why: Several new models/ files used as base_model targets only set limit.context and omit limit.output (magistral-small-2509, all three ministral-3-*-instruct-2512, both voxtral-*-2507, openai/gpt-oss-safeguard-20b). AGENTS.md requires complete lab entries when hosts inherit them; host-only output overrides leave lab metadata incomplete and make inheritance fragile. Action: Add verified limit.output (and any other missing required lab fields) on every new lab TOML before relying on Bedrock overrides.
• [high] [possible mistake] providers/amazon-bedrock/models/anthropic.claude-opus-4-6-v1.toml:6 - Check: Bare/unprefixed Bedrock IDs should carry the correct source-region Standard rates, not a mismatched geo. Why: Bare anthropic.claude-opus-4-6-v1 and anthropic.claude-sonnet-4-6 are re-priced to the 1.1× band ($5.50/$27.50 and $3.30/$16.50) with eu-west-2 price-list citations, while the Global twins stay on the $5/$25 and $3/$15 band. That split is internally inconsistent unless the bare IDs are truly EU-only on-demand. Action: Confirm the bare model IDs’ actual source-region Standard dimensions; if they are US/global on-demand, restore US list rates and keep 1.1× only on AU/EU/JP/US geo profiles.
• [medium] [possible mistake] providers/amazon-bedrock/models/mistral.ministral-3-14b-instruct.toml:1 - Check: Provider files must override lab modalities/attachment when the host surface differs. Why: Switching these entries to base_model = "mistral/ministral-3-*-instruct-2512" and dropping local [modalities]/attachment makes Ministral 14B and 8B inherit lab attachment = true and input = ["text", "image"]. The previous Bedrock definitions were text-only / attachment = false (3B already kept image). Action: If Bedrock Converse for 14B/8B is text-only, restore explicit attachment = false and text-only [modalities] overrides (or document host multimodal support with a first-party Bedrock citation).
• [medium] [possible mistake] providers/amazon-bedrock/models/google.gemma-3-27b-it.toml:8 - Check: tool_call overrides must match the Bedrock host, not only the lab default. Why: Lab google/gemma-3-*-it sets tool_call = false, but Bedrock keeps/sets tool_call = true on Gemma 3 27B and 4B while 12B inherits false. That split is suspicious without host evidence. Action: Align tool_call across the Gemma 3 Bedrock trio to whatever Converse actually supports, and cite the Bedrock model card/API behavior.
• [low] [possible mistake] providers/amazon-bedrock/models/deepseek.r1-v1:0.toml:8 - Check: Capability overrides need host evidence when they flip lab defaults. Why: tool_call = false is newly forced on bare and US DeepSeek R1 while lab deepseek/deepseek-r1 is tool_call = true. Comments cover always-on reasoning and output caps, not tools. Action: Verify Bedrock R1 tool calling; keep the override only with a card/API note, otherwise drop it so the lab value inherits.

[rekram1-node]

Addressed the five review items in 8ffef3a, with additional primary-source checks and direct Converse probes:

1. Complete new labs — fixed. All 11 new lab entries now contain every required field, including both context and output. Added the seven missing outputs: Magistral 131,072; Ministral 262,144; Voxtral 32,768; Safeguard 20B 131,072. Headers cite exact native configs/examples. For open-weight models without a separately published completion cap, these are explicitly documented shared-window generation ceilings, not invented hosted-API maxima. Bedrock's smaller output overrides remain separate. Voxtral context was also made precise at 32,768 from native params.json.

2. Bare Claude pricing — confirmed, retained. Both bare IDs succeed in London. Fresh us-east-1 calls reject both with HTTP 400: “on-demand throughput isn't supported … use … an inference profile.” The London Standard dimensions are:

   • Opus 4.6: $5.50/$27.50, SKUs 49XYAK32S4QDU7VP / K7N8ZERVSHSFXCVC.
   • Sonnet 4.6: $3.30/$16.50, SKUs ZMG85E5S2RJPKJCH / CH7Y686BV8D7WX7K.
   • The cheaper Global twins use separate _Global dimensions. Their price difference is expected, not a US/EU mismatch.
   • London source prices, Opus card, Sonnet card.

3. Ministral 8B/14B vision — confirmed, retained. Native Converse accepted generated solid-color PNGs and correctly answered Blue (8B) and Yellow (14B), HTTP 200. The 8B and 14B cards affirm Image input and Converse. Added live-evidence comments.

4. Gemma 3 tools — aligned false across 4B/12B/27B. All three can emit genuine toolUse blocks under forced ANY selection, but all three reject valid native toolResult continuations with a backend role-alternation error. An independent retry adding assistant text also failed. Removed the positive overrides: an agent-facing tool capability should not promise a complete tool-use loop when its result continuation is broken. Comments explicitly preserve the distinction between forced call generation and the failed roundtrip.

5. DeepSeek R1 tools — false confirmed. us.deepseek.r1-v1:0 explicitly rejects toolConfig with “This model doesn't support tool use.” Confirmed using AUTO and with toolChoice omitted, so this is not merely a forced-choice restriction. Both retained R1 overrides now cite that exact result and the AWS parameter guide.

Combined bun validate, explicit new-lab completeness checks, and git diff --check pass. Generated-catalog comparison still reports zero existing model removals, route changes, or lifecycle-status changes.

[github-actions]

Action items

• [high] [possible mistake] providers/amazon-bedrock/models/jp.anthropic.claude-opus-5.toml:6 - Check: Regional Claude geo pricing consistency (JP 1.1x vs US/global 1.0x). Why: This PR updates JP Opus 4.7/4.8, Haiku 4.5, Sonnet 4.5, and Sonnet 4.6 to ap-northeast-1 1.1x rates (5.50/27.50, etc.), and US/AU Sonnet 5 / Opus 5 follow the same geo pattern, but JP Opus 5 stays at 5/25 and JP Sonnet 5 stays at 2/10. That understates JP on-demand cost for those two profiles relative to the PR’s own pricing rule. Action: Align jp.anthropic.claude-opus-5 and jp.anthropic.claude-sonnet-5 costs (and cache rates) with the same JP source-region dimensions used for the other JP Claude entries, or document why these two SKUs are intentionally 1.0x.
• [medium] [possible mistake] models/meta/llama-4-scout-17b-instruct.toml:14 - Check: Host-specific context limits after lab metadata changes. Why: Lab context is raised 3_500_000 → 10_000_000, and Bedrock bare/US Scout files only override limit.output (to 8_192), so resolved Bedrock context becomes 10M. Other hosts already cap Scout far lower (often ~128k–1.3M). The Bedrock edits cite live maxTokens checks, not a 10M context verification. Action: Confirm the AWS Bedrock Scout context window; if Bedrock is below 10M, add an explicit Bedrock [limit].context override on meta.llama4-scout-17b-instruct-v1:0 and us.meta.llama4-scout-17b-instruct-v1:0 (and keep the 10M value only on the lab entry if that is the native Meta figure).
• [medium] [possible mistake] providers/amazon-bedrock/models/moonshotai.kimi-k2.5.toml (temperature line in diff) - Check: Override-only base_model files must not invent capability deltas. Why: Shared lab models/moonshotai/kimi-k2.5.toml has temperature = false, but the Bedrock entry keeps temperature = true after switching to base_model, with no Bedrock wire/live note for temperature (unlike the reasoning toggle / maxTokens notes). Action: Drop temperature so it inherits false, or keep the override only with a leading comment citing Bedrock acceptance of temperature on this ID.

[rekram1-node]

Superseded by smaller, independently validated PRs based on current dev:

• fix(amazon-bedrock): correct Anthropic catalog #6790 — Anthropic
• fix(amazon-bedrock): correct Meta, Mistral, and Writer catalog #6791 — Nova
• fix(amazon-bedrock): correct OpenAI catalog #6792 — OpenAI
• fix(amazon-bedrock): correct Nova catalog #6793 — Meta / Mistral / Writer
• fix(amazon-bedrock): correct open model catalog #6794 — remaining open models

The latest reviewer findings are included in these splits: Tokyo Claude 5 Standard pricing is corrected, Scout’s inherited 10M Bedrock context is cited, and Kimi K2.5’s temperature override has live Converse evidence. No existing model removals are included.

[rekram1-node]

Correction to the split mapping above: #6791 is Meta / Mistral / Writer, and #6793 is Nova. The other three mappings are unchanged.