fix(gcp-to-aws): record Fable 5.1 pricing and enforce restricted-model selection - #267
Conversation
![kb-autoupdate[bot]](https://github.githubassets.com/images/gravatars/gravatar-user-420.png){kind=avatar}
Source: https://aws.amazon.com/about-aws/whats-new/2026/09/claude-fable-5-1-aws/ Proposed by the knowledge auto-update pipeline; every edit's justification is in the PR body.
…e Mythos row Reworks the pipeline proposal. Verified 2026-09-03 against the Bedrock pricing page (Anthropic tab), the Fable 5.1 model card, and the Sep 1 What's New posts: - Fable 5.1: anthropic.claude-fable-5-1, GA Sep 1, 2026, 1M context, $10/$50, cache write 12.50 (5m) / 20.00 (1h), cache read 0.25 (0.025x — the one rate that differs from Fable 5's 1.00). Batch not listed on the Bedrock page. Covered Model: requires the aws_review data-retention mode; commercial regions are us./global. CRIS only (no in-region); GovCloud (US) has in-region access. - Mythos 5.1 row removed: it is a gated Preview for vetted cyber/bio research organizations, requires provider_data_share, and is not a migration target. Mentioned once in prose so readers know why it is absent. - ai-anthropic-to-bedrock: "do not default to Fable 5" now covers 5.1.
leon1418
changed the title
944e89c
… widen the _unverified_ gate Follow-up to the review on awslabs#267: - estimate-ai.md: the "_unverified_ is blocking" rule was scoped by its heading to OpenAI models, so an Anthropic row with an unverified price passed every guard. It now applies to every provider; the Price List API gap is given as the usual cause rather than the scope. - pricing-cache.md and ai-anthropic-to-bedrock.md: "do not default" now names the Fable / Mythos frontier line (5, 5.1, and successors) instead of one version string, so a 5.2 launch does not silently become recommendable. - agent-advisor anthropic catalog disclaimer: names Fable 5.1 / Mythos 5.1, states each model's data-retention opt-in mode (provider_data_share vs aws_review) and Fable 5.1's CRIS-only footprint.
…ine, not one version (review on awslabs#267)
leon1418
left a comment
leon1418
left a comment
There was a problem hiding this comment.
[🤖 AI review 🤖] Review of PR #267 at head efacaf54 (8 files, +14/-10)
Summary: Records Claude Fable 5.1 (GA Sep 1, 2026) in the pricing cache with verified pricing ($10/$50, cache read $0.25), access model (Covered Model, aws_review opt-in, CRIS-only in commercial regions), and model ID. Removes the pipeline-proposed Mythos 5.1 row (gated Preview, not a migration target). Widens the unverified-is-blocking pricing guard from OpenAI-only to all providers. Makes the "do not default" guardrails version-agnostic (Fable/Mythos line, not just Fable 5). Updates the agent-advisor JSON catalog disclaimer to name each model data-retention opt-in mode.
Findings:
-
Mandatory (1): The pricing-cache.md narrative at L386 says "Both are Anthropic Covered Models: using them requires opting the account into the aws_review data-retention mode" -- but Fable 5 requires provider_data_share, not aws_review. The JSON disclaimer in the same PR correctly distinguishes the two. Risk: a reader following the narrative opts into aws_review only and gets AccessDenied on Fable 5 invoke.
-
Nit (1): Sonnet 5 quick-reference row still shows $2/$10 intro pricing that expired Aug 31. Pre-existing; noted since Last updated is bumped to Sep 3.
Cross-plugin parity: PASS -- all 4 file pairs (advisor / migrate) share identical blob SHAs. 0 drift allowlist changes.
Existing threads: ayn-builds raised 3 findings on the pipeline original proposal; all 3 are addressed in commits 944e89c (verified pricing, Mythos row removed) and efacaf5 (version-agnostic guardrails, provider-scoped caveat). Author replies reference specific commits.
CI: 8/8 SUCCESS (build, gitleaks, bandit, semgrep, checkov x2, Bandit, gitleaks).
Approvals: 0 (2 prior reviews DISMISSED by new commits).
Mergeable: MERGEABLE after the mandatory finding is fixed and approvals obtained.
Validation evidence: Diff read via gh pr diff, file contents verified at head via GitHub Contents API, blob SHA parity confirmed for all 4 cross-plugin pairs. No tests executed by this review (change is reference data only; PR description reports drift:check OK, fmt:check OK, node tests 62+62+12).
…are — align the catalog with the model cards The anthropic catalog disclaimer (written 2026-07-21) said Fable 5 needs the provider_data_share retention mode, and efacaf5 carried that forward. The Fable 5 model card, read 2026-09-03, says: "To use this model, you must opt in to AWS review by setting your data retention mode to aws_review". The same sentence appears on the Fable 5.1 card. provider_data_share is what the Mythos 5 and Mythos 5.1 cards require. The pricing-cache narrative ("both are Covered Models … aws_review") was already correct; the catalog was the stale side.
|
Re the mandatory finding at head efacaf5: the narrative was right and the JSON catalog was wrong, per the live model cards — Fable 5 and Fable 5.1 both require |
…d make the guards machine-readable Second review round on awslabs#267 (ayn-builds). All five findings verified against the source files and the Bedrock pricing page before changing anything. - New Status value `restricted (<reason>)` in ai-model-lifecycle.md (Status table, Estimate rules, mapping-guide rules, selection steps). `active` is defined as "no restrictions", so `active (Covered Model; …)` passed every downstream gate; the estimate checklist screened Legacy only. estimate-ai.md gets the matching lifecycle bullet and a checklist item. Claude Fable 5 and Fable 5.1 rows now carry `restricted (Covered Model; aws_review opt-in …)`. - design-ai.md: a provider-agnostic "restricted models are never recommended_model / backup_model" rule at the top of Part 1 — the loader routes by ai_source, so a rule living only in the Anthropic guide and the pricing cache never reached Gemini- or OpenAI-source designs. The gdpr row now excludes candidates with no `eu.` Geo profile (Fable 5 / 5.1 have only `us.` / `global.`), and the soc2/ccpa row names the aws_review / provider_data_share retention conflict. - pricing-cache.md: Fable 5.1's Model ID is the invokable `us.` form (the bare id is not accepted on bedrock-runtime; no in-region form exists). Batch for Fable 5 AND Fable 5.1 is `N/A`: the Bedrock page hard-codes N/A in both the Global and the Geo/In-region Anthropic tables (read 2026-09-04) — the 5.00 / 25.00 Fable 5 batch values were Anthropic first-party rates, not Bedrock. Footnote added; "only rate difference" wording corrected. - Catalog disclaimer: revision history removed (belongs in git), dated model-card evidence clause added, Mythos characterized as a gated preview that no account-level opt-in reaches.
|
Second round addressed in 371bc12 (both plugin copies; drift:check, fmt:check, node tests 62+62+12, agent-advisor pytest 322 all green locally). Structural changes: a new |
…cycle page cannot produce it The refresh instruction told the next cache update to rewrite the whole Status column from the Bedrock model lifecycle page, which publishes only active / legacy / EOL. Followed literally (and the kb-autoupdate pipeline follows it literally), that resets Fable 5 / 5.1 to `active` and silently disables every guard that keys on `restricted`. The instruction now refreshes only the lifecycle-derived values and keeps `restricted` unless the model card's access requirement itself changed.
leon1418
changed the title
…w the standard price Anthropic's pricing page states the $2/$10 introductory rate for Claude Sonnet 5 "is now the standard price" and that "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur." The Bedrock pricing page (Anthropic tab, read 2026-09-03) lists Sonnet 5 at $2.00/$10.00 with no promotional qualifier. Our documents had been pricing Sonnet 5 at the cancelled $3/$15 "steady-state" rate in every comparison, overstating it by a third. Every Sonnet 5 comparison is re-derived at $2/$10 (blended 2:1 in:out, per the house rule), and several conclusions flip: - GPT-5.6 Terra vs Sonnet 5: "Terra 19% cheaper" -> Sonnet 20% cheaper - GPT-5.4 vs Sonnet 5: "near parity" -> Sonnet 36% cheaper - GPT-5.5 vs Sonnet 5: 52% -> 68% cheaper - GPT-4o vs Sonnet 5: "Sonnet 29% more expensive" -> Sonnet 7% cheaper - GPT-5.2 vs Sonnet 5: "source 17% cheaper" -> Sonnet 20% cheaper - GPT-5.1/5 and GPT-4.1: source advantage shrinks to 11% / 14% - GPT-4 Turbo / GPT-4: 58% / 82% -> 72% / 88% cheaper - Gemini 3.1 Pro vs Sonnet 5: "+24%" -> -13%; Sonnet 5 is now the cheaper side Files: pricing-cache (narrative + quick-reference status), clarify-ai and clarify-ai-only baseline tables, ai-openai/ai-gemini/ai-anthropic design refs, and the test comment in test_bedrock_pricing.py (the STATIC_FALLBACK already carried 0.002/0.010). The "do not default to Fable 5" sentences in the openai and gemini refs now also name Fable 5.1 (see #267). Both plugin copies updated.
Preserve Fable restricted statuses and N/A batch rates alongside upstream Sonnet 5 and Opus 4.8 unverified batch markers. Align the estimate guide examples with the merged pricing table.
4 participants
The Fable 5.1 launch introduced a new Bedrock model with access and regional restrictions that must be reflected in both pricing data and model selection. This PR records verified rates, excludes restricted frontier models from default recommendations, and keeps those restrictions intact when the cache is refreshed.
Changes
us.anthropic.claude-fable-5-1for US Geo inference) with 1M context, $10/$50 per million input/output tokens, cache writes of $12.50/$20 for 5m/1h, and $0.25 cache reads. Both Fable 5 and 5.1 haveN/ABedrock batch rates.restricted: both require theaws_reviewdata-retention opt-in. Exclude restricted models fromrecommended_model,backup_model, and mapping defaults; describe them only on explicit request with their access requirements. Keep Mythos out of the recommendation table because it is a gated preview, not unlocked by an account-level opt-in alone._unverified_pricing gate to all providers, and preserverestrictedduring lifecycle-only cache refreshes.us./global.inference profiles onbedrock-runtime; Fable 5 also has in-region access throughbedrock-mantlein us-east-1. Neither has aneu.Geo profile, so the existing EU filter remains. Generic disclosures now name any applicable endpoint-specific regional or CRIS-only restrictions.N/A. Keep the estimate guide's unverified examples consistent with those rows. The Sonnet 5 pricing and comparison corrections from merged fix(gcp-to-aws): align Sonnet 5 pricing, guidance, and report estimates #269 are included through main.All six changed file pairs are synchronized between the advisor and migrate plugins. #268 is superseded by this PR.
Sources
Validation
dprint check --incremental=false: passed.git diff --check: passed.This branch includes main at
57708ee0, including merged #269. The September 16 merge resolves the two pricing-cache conflicts, preserving the Fable/restricted-model additions and upstream batch-price warnings. Both estimate-guide copies also distinguish the remaining unverified rates from Fable'sN/Arates.