* feat(bedrock): serve the OpenAI models on bedrock-runtime's native Responses API
AWS serves the OpenAI models on bedrock-runtime through an OpenAI-compatible
surface at /openai/v1/responses, alongside Converse. LiteLLM had no Responses
config for the bedrock provider, so /v1/responses fell back to the Chat
Completions bridge and was translated into Converse. A realistic Codex session
does not survive that translation: its function_call / function_call_output
history becomes Converse toolUse / toolResult blocks with no toolConfig, and
Converse rejects the request outright.
Add a Responses config for that surface, opted into per model from the price-map
supported_endpoints so models without the signal keep the bridge exactly as
before. Auth is Bearer when a Bedrock API key is present, SigV4 otherwise.
Both Bedrock endpoints reject the Codex history item types agent_message,
context_compaction and local_shell_call, so the normalization bedrock_mantle
carried privately moves into a shared module and both providers use it. They are
history items, so they only bite from the second turn onward -- a first-turn
smoke test passes and hides the problem. Verified against bedrock-runtime with
global.openai.gpt-5.6-sol: additional_tools is accepted there (unlike on
bedrock-mantle) while those three types are rejected, so the two endpoints do
not share one validator and each provider opts in explicitly.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* fix(bedrock): build the Responses endpoint from the region's partition suffix
get_complete_url hardcoded amazonaws.com in an f-string, so every non-commercial
partition got the wrong host: cn-north-1 resolved to amazonaws.com instead of
amazonaws.com.cn, and GovCloud/ISO regions were wrong the same way. Defer to
BaseAWSLLM._select_default_endpoint_url, which this config already inherits and
which resolves the suffix per partition.
test_no_fstring_hardcodes_the_commercial_dns_suffix scans the whole tree, so it
caught this even though it is not one of this PR's test files. Register the
config in ENDPOINT_BUILDERS so the cn/GovCloud endpoint sweep covers this
surface from now on rather than only the f-string guard.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* feat(bedrock): opt the gpt-6 family into the native Responses API
* fix(bedrock): drop the Responses tool types bedrock-runtime rejects
Codex sends a web_search tool on every turn. api.openai.com runs that tool
itself, and the Converse bridge dropped it silently, but bedrock-runtime's
native Responses endpoint rejects the whole request with 400 "web search is
not supported for this request". Filter the request's tools down to the
types bedrock-runtime's own validation error names, logging what was dropped,
through a helper shared with the Mantle route, which already did the same.
* fix(bedrock): emulate file_search and collapse custom Responses paths
* fix(bedrock): keep background and remote image inputs working on the native Responses route
* fix(bedrock): inline remote images inside tool outputs on the native Responses route
* fix(bedrock): inline remote computer screenshots on the native Responses route
---------
Co-authored-by: Leonardo Freitas dos Santos <leonardo.freitas.s@outlook.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(prices): add baseten/zai-org/GLM-5.3-Fast pricing with cost tracking e2e
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(e2e): assert message instead of comment on breakdown row
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(baseten): drop the live e2e cost tracking test
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: kerry <kerry@berri.ai>
* feat(cost-map): add Azure Foundry pricing for gpt-6-sol and gpt-6-luna
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost-map): give azure/eu gpt-6-sol and gpt-6-luna full model metadata
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(pricing): correct cached-token fields on realtime cost-map entries
azure/gpt-realtime-2 was the only member of the gpt-realtime-2 family priced
on one side of its cached-audio meter. Azure publishes that meter as
"gpt-realtime-2 Audio cd inp Gl 1M Tokens" at 0.4 per 1M and charges the
same rate for the write that populates the cache and the read that hits it,
so cache_creation_input_audio_token_cost lands at 4e-07, matching
azure/gpt-realtime-2.1, azure/gpt-realtime-2.1-mini and the openai
gpt-realtime-2 entry. No cost path reads that field yet, so this corrects
what get_model_info reports rather than what anything bills.
The gemini Live entries go the other way. Google's Vertex context-caching
page publishes separate supported-model lists for implicit and explicit
caching, and no Live or native-audio model is in either one. Its pricing
page prints N/A in both cached-input columns for every Gemini 2.5 Flash
Live API row, where plain 2.5 Flash and 2.5 Flash-Lite both carry real
cached prices, and the Vertex model card for the family marks context
caching not supported outright. Vertex never reports cachedContentTokenCount
on a Live session either, including for a byte-identical 7,021-token prefix
replayed across sessions minutes apart, which is well past the 2,048-token
minimum the same page sets for the Gemini 2 family.
So the 7.5e-08 on the two preview siblings priced something the provider does
not sell, and supports_prompt_caching on all three claimed a capability the
model does not have. The rate comes out. The flag is set to false rather than
removed, because get_model_info maps an absent key to None, and None is how
this map spells "nobody checked" across the 2,788 entries that omit it, where
false records the vendor's documented no. Both readers of the flag gate on
`is True`, so nothing bills or behaves differently either way.
Only the cached fields change on the two 09-2025 preview entries. Their
source field points at the Gemini API pricing page rather than the Vertex
one, so they describe a different surface with its own published limits, and
their context windows are left alone rather than assumed to match the Vertex
model card that drives the GA entry.
Tests cover all three halves: the family invariant that a cached audio read
implies an equal cached audio write, a cached count on a Live entry leaving
the bill at the fresh-input total instead of adding the old 7.5e-08, and
supports_prompt_caching answering false for all three entries while still
answering true for 2.5 Flash, so the false cannot be a swallowed lookup
error.
* fix(cost): correct gemini-live-2.5-flash-native-audio limits and capabilities
Google's model card for model ID gemini-live-2.5-flash-native-audio gives a
128K context window and 64K maximum output tokens, and marks structured
output, context caching and URL context as not supported. Its modality list
is text in and out, image in, audio in and out, and video in, with no
document input of any kind.
The entry advertised a 1M context window, an off-by-one 65535 output cap, and
three capability flags the vendor marks unsupported. Context caching is the
fourth and is handled in the cached-fields change alongside its two preview
siblings.
Both the bare id and vertex_ai/gemini-live-2.5-flash-native-audio resolve to
this single entry, so the test drives the corrected values through both.
* test(integration): cover live preview cached tokens billed at the fresh rate
Co-authored-by: Marty Sullivan <marty@martysullivan.com>
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(cost): cite dated sources for Live entry pins and drop restating docstrings
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Marty Sullivan <marty@martysullivan.com>
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): send json_schema as a forced tool on Claude Opus 4.7 and 4.8 Converse
Bedrock rejects outputConfig.textFormat on Opus 4.7 and 4.8 with
"output_config.format: Extra inputs are not permitted", and the AWS
model cards list structured outputs as not supported for both, so
their cost-map entries no longer claim supports_native_structured_output
and json_schema requests fall back to the json_tool_call tool.
Fixes#27846
* test(bedrock): assert Opus 4.7 and 4.8 inline the schema on Invoke, move the native case to Sonnet 4.6
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix: repair seven regressions caught by CircleCI on main
- vertex_ai: stop treating fine-tuned endpoint ids (numeric or
vertex_ai/gemini/<id>) and gemma models as Gemini 3+, which injected
temperature=1.0 and Gemini 3 thinking config into their requests (#42465)
- cost: price Azure DALL-E 3 from its azure/<quality>/<size>/dall-e-3 rows;
it only worked through the OpenAI rows that #42435 removed
- bedrock: stream bedrock/invoke/moonshot through an OpenAI-shaped chunk
decoder; the generic decoder dropped every chunk, which the
supports_response_schema flag from #42338 un-skipped in CI
- proxy: keep the public model_group on pre-routing rejections so the Usage
page groups them under the model name, not the deployment (#41077)
- cost map: mirror the base rows' capability flags onto Bedrock regional and
cross-region copies (#42254 and later syncs)
- whitelist the new regional Bedrock rows from #42543 and #42588 for the
converse routing check, following the existing regional-row convention
* fix(model-prices): mirror capability flags onto ap-southeast-3 bedrock rows
* refactor(bedrock): tighten types on the moonshot stream decoder and its tests
* chore(cost-map): remove models past their deprecation date
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(cost-calc): drop the empty parametrize left behind by the gemini web search removal
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost-map): drop merge base block left by conflict resolution
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(cost-calc): drop gemini image cost tests pinned on removed model
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(model_prices): registry audit 2026-09-22, absorb open pricing PRs
Rolls the open registry-only PRs into one PR after re-verifying every value against the official provider source: OpenAI, Azure, Vertex AI and Gemini batch cache-read prices, Baseten model metadata from the authenticated inference API, Bedrock eu-west-2 Nemotron Super 3 pricing from the AWS offer file, and OpenRouter prices refreshed from the live OpenRouter models API
Co-authored-by: sinksilk <785976238@qq.com>
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(model_prices): add groq/llama-guard-3-8b from the Groq model page
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(model_prices): refresh openrouter deepseek aliases from live api and drop stale off-peak windows
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(model_prices): resolve baseten merge conflicts against main
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
---------
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: sinksilk <785976238@qq.com>
Add gpt-6-sol and gpt-6-luna to the model cost map with pricing from the OpenAI pricing page and reasoning effort levels none through max. Extend the long-context priority pricing and reasoning effort capability tests to cover both models
Co-authored-by: kerry <kerry@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost): bill batch prompts above 272K at OpenAI's long-context batch tier
* fix(cost): mirror batch long-context keys on custom pricing params
Register the two *_above_272k_tokens_batches keys on CustomPricingLiteLLMParams so a per-deployment override stays out of the shared backend key, add them to the inline model-info schema and alias-count tests, and build LiteLLM_Params and GenericLiteLLMParams through model_validate at the two dict-splat call sites so basedpyright's reportArgumentType budget ratchets down instead of blocking the new fields.
* fix(cost): add the gpt-5.5-pro batch long-context tier and ignore malformed batch tier keys
* fix(cost): bill cached batch tokens at OpenAI's cached batch rate
Adds cache_read_input_token_cost_batches and
cache_read_input_token_cost_above_272k_tokens_batches for the tiered
OpenAI entries at half the standard cached rate, bills cached batch
tokens at that rate per output line, and parses string-valued batch
rates in deployment-level model_info.
* fix(cost): bill batch cache writes at the batch cache-write rate and carry published batch rates for one-sided deployments
OpenAI's Batch table prices cache writes for gpt-6-astra, gpt-5.6, gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna at half the standard cache-write rate, so the cost map gains cache_creation_input_token_cost_batches and its above_272k tier for those entries and batch cost pulls written tokens out of the input bucket at that rate; models without the key keep billing writes at the batch input rate.
A deployment declaring only one side of its batch pricing now carries every published batch rate of the other side (tier, cached, cache write), its own keys win, and a lone tier, cached or cache-write batch key counts as declared pricing instead of being ignored.
* fix(cost): select the batch long-context tier from any batch tier key
A deployment that declares its own flat standard input rate keeps every
published batch rate of the output direction, including the 272K output
tier, but the tier was only ever selected when an input tier key was also
present. Detect the crossed tier from any of the four batch tier keys so
the carried output, cache-read, and cache-write tiers bill at their tier
rate above 272K tokens.
* chore(proxy): keep the OpenAPI snapshot as CI generates it
* fix(cost): pick each batch price component's tier from its own keys
The batch rate picker crossed one threshold for every component, so a
deployment declaring only an output tier also moved its input, cached, and
cache-write rates to that cutoff. Each component now crosses its own
*_above_<N>k_tokens_batches keys and falls back to its flat key.
The JSON schema is regenerated with the generator as it is on main:
cost-map-guard renders the PR's cost map with the base branch's generator,
so the descriptions for the new batch cache keys move to a follow-up.
* chore(proxy): restore the lazy OpenAPI snapshot to what CI's Python 3.12 generates
The merge commit carried a snapshot regenerated on a Python 3.14 venv, which dedents
docstrings at compile time, so one description line differed from the file CI regenerates
on 3.12 and the schema.d.ts sync check went red. The snapshot is byte-identical to main again
