Skip to content

RubyLLM2 W2-04: Port the Claude 5.5 and GPT-6 Bedrock registry entries to the 2.0 fork - #40

Merged
samuelboland merged 1 commit into
supernova-2-0from
dux-rubyllm2-w2-04-registry-port
Nov 10, 2026
Merged

samuelboland merged 1 commit into
supernova-2-0from
dux-rubyllm2-w2-04-registry-port

Conversation

@staging-supernova-dx-appf-io

@staging-supernova-dx-appf-io staging-supernova-dx-appf-io Bot commented Nov 10, 2026 •

Copy link
Copy Markdown

What this does

Adds the bedrock registry entries that lib/ruby_llm/models.json on supernova-2-0 was missing for models Supernova runs in production: Claude Sonnet 5.5, Opus 5.5 and Haiku 5.5 (bare id plus the au., eu., global., jp. and us. prefixes upstream has), and GPT-6 Sol, GPT-6 Luna, GPT-6.1 Sol and GPT-6.1 Terra (bare, us. and global.). Without them RubyLLM 2.0 cannot find these ids, so they have no pricing, capabilities or context window, and Supernova's thinking clamp returns 'none' for them.

29 entries, each inserted at the position the file's existing bedrock order puts it. The file is otherwise unchanged: a line diff against the base is 2,621 insertions and 8 deletions (the 8 are the Haiku 4.5 price lines below). GitHub's larger deletion count comes from how its diff lines up the inserted blocks.

  • 26 entries are upstream main's, copied as is, except for the Haiku 5.5 change below.
  • The three GPT-6.1 Terra entries are not upstream's (upstream has no Terra entry). They are written from AppFolio's Bedrock agreement offer (2026-11-09), in $ per million tokens:
    • openai. and us.: 1.1 / 5.5 / 0.055 / 1.375, and 2.2 / 8.25 / 0.11 / 2.75 above 272,000 input tokens.
    • global.: 1 / 5 / 0.05 / 1.25, and 2 / 7.5 / 0.1 / 2.5 above 272,000.
    • Effort values are none through max. Terra accepts none (checked live 2026-11-09); GPT-6.1 Sol does not.
    • pdf input is listed on the bare id only, where it was checked live.

Differences from upstream

  • Haiku 5.5, no thinking toggle. The six Haiku 5.5 entries have no {"type":"toggle"} in metadata.reasoning_options, only effort (low to max). Upstream has the toggle; Sam decided on 14 Oct that Haiku 5.5 gets no thinking-off toggle.
  • Haiku 4.5 us. price. The existing entry us.anthropic.claude-haiku-4-5-20251001-v1:0 changes from the regional rate on supernova-2-0 (1.1 / 5.5 / 0.11 / 1.375) to the global rate (1 / 5 / 0.1 / 1.25), in both pricing.text_tokens.standard and metadata.cost. It bills at about 94% of the regional rate, and the 1.16 fork has carried the global price since fork PR fix: correct Bedrock us. regional premium & cache-read rates #34. No other field or Haiku 4.5 entry changes.

Not changed

openai.gpt-5.6-sol is left as it is on supernova-2-0: 4.4 / 22 / 0.44 / 5.5, with a 272,000 tier at 8.8 / 33 / 0.88 / 11. The 1.16 fork carries 5 / 30 / 0.5 / 6.25 with no tier, and that was wrong: the October 2026 AWS bill charged exactly 4.40 / 22.00 / 0.44 / 5.50. The new spec asserts the 2.0 values, so a later change to them is deliberate.

No metadata.endpoint marks are added to the GPT entries; RubyLLM2 W2-02 adds them.

Tests

New spec/ruby_llm/providers/bedrock/registry_prices_spec.rb (25 examples), reading the bundled registry through RubyLLM.models.find(id, provider: 'bedrock'):

  • Exact standard rates for the four Supernova-pinned Claude ids (Haiku 4.5 us., Opus 5.5 us., Sonnet 5.5 us., Haiku 5.5 us.).
  • Haiku 5.5 on us. and global.: long-context rates and the 100,000 threshold, both from the pricing hash and rebuilt from metadata.cost.tiers alone.
  • Every Haiku 5.5 entry: reasoning_options is exactly effort low to max, with no toggle.
  • Opus 5.5 us., Sonnet 5.5 us., GPT-6 Sol, GPT-6 Luna, GPT-6.1 Sol and GPT-6.1 Terra resolve on bedrock with their context windows and max output tokens.
  • GPT-6.1 Terra (bare and us.): the US rates, the 272,000 tier, and effort none through max.
  • openai.gpt-5.6-sol keeps its billed rates.

Runs (with LANG=C.UTF-8; the container's default US-ASCII locale cannot read models.json):

  • New spec, seeds 1234, 48211 and 9071: 25 examples, 0 failures each time.
  • spec/ruby_llm/models_json_validation_spec.rb, spec/ruby_llm/providers/bedrock/models_spec.rb and spec/ruby_llm/providers/bedrock/mantle_spec.rb together: 70 examples, 0 failures, 1 pending (the existing mantle Claude example that needs an AWS agreement).
  • bundle exec rubocop on the new spec: no offenses.

The whole suite was not run. No live calls or cassettes were needed.

This PR targets supernova-2-0, not main, which stays the 1.16 rollback line. It is not for upstream crmne/ruby_llm. Supernova moves to this branch later, in RubyLLM2 W3-01.

The 2.0 registry has no bedrock entry for Claude Sonnet, Opus or Haiku
5.5, or for GPT-6 Sol, GPT-6 Luna, GPT-6.1 Sol or GPT-6.1 Terra, so
Supernova cannot find the models it uses in production: no pricing,
no capabilities and no context window.

This ports upstream main's 26 entries for those models, without the
thinking-off toggle on Haiku 5.5, and adds three GPT-6.1 Terra entries
written from AppFolio's Bedrock agreement offer. The us. Haiku 4.5
entry moves to the global rate, which is what that profile bills and
what the 1.16 fork already carries. openai.gpt-5.6-sol keeps its 2.0
rates, which match the October 2026 AWS bill; the spec pins them.

Co-authored-by: Sam Boland <sam.boland@appfolio.com>

@samuelboland samuelboland left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Verified: the 29 entries in models.json match the brief exactly and the only other change is the Haiku 4.5 us. price; spec passes under three seeds; validation, models and mantle specs green.

@samuelboland
samuelboland merged commit af91288 into supernova-2-0 Nov 10, 2026
1 check failed
@staging-supernova-dx-appf-io

Copy link
Copy Markdown
Author

🤖 Supernova Code Review — View trace

🔁 Re-run this review

— via automation "Code Review Opus 5.5 Medium" (run), installed by Sam Boland (sam.boland@appfolio.com)

@staging-supernova-dx-appf-io staging-supernova-dx-appf-io Bot left a comment

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Reviewed the new registry_prices_spec.rb and spot-checked the added and changed models.json entries at the PR head: us.anthropic.claude-haiku-4-5-20251001-v1:0 (L19107), us.anthropic.claude-haiku-5-5 (L19200), us.anthropic.claude-opus-5-5 (L19900), openai.gpt-6-luna / gpt-6-sol / gpt-6.1-sol / gpt-6.1-terra (L16886–17271), and global.openai.gpt-6.1-sol / gpt-6.1-terra (L12796–12986). The prices, tiers, thresholds, effort values, and the missing Haiku 5.5 toggle all match what the PR description says. I also checked that Model#to_h returns fresh pricing hashes, so the metadata.cost.tiers rebuild example doesn't change the shared registry model.

The only finding is one nit, inline: the global Terra entry has no test.

🔁 Re-run this review

— via automation "Code Review Opus 5.5 Medium" (run), installed by Sam Boland (sam.boland@appfolio.com)

%w[openai.gpt-6.1-terra us.openai.gpt-6.1-terra].each do |id|
it "prices #{id} at the US Terra rates with a 272,000 tier" do
text = bedrock_model(id).pricing.text_tokens

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

nit: global.openai.gpt-6.1-terra (models.json L12892) is also hand-written from the agreement offer (1 / 5 / 0.05 / 1.25, tier 2 / 7.5 / 0.1 / 2.5), but no spec checks it. Adding a matching example for the global rates would catch a typo in the only Terra entry this list leaves out.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants