* feat(bedrock): send grok chat completions through runtime openai path
Unspecified bedrock grok was rewritten to Converse. Chat completions now hit bedrock-runtime /openai/v1/chat/completions, and converse/ still uses Converse
* feat(bedrock): serve gpt-oss and gpt-5.6 chat completions on runtime's native openai path
* fix(bedrock): route gpt-oss response_format to Converse and decide the route once from the raw request
* fix(bedrock): serve region-path and GovCloud gpt-oss ids on native Chat Completions
The cost-map parity tests require every regional variant of a flagged id to carry the same supports_ flags, so the six us-gov gpt-oss entries now carry the native-route flags too. A region path in the model name (bedrock/us-gov-west-1/openai.gpt-oss-20b-1:0) is routing, not a different model: the route is looked up on the id after the path, the path's region picks the endpoint and the SigV4 scope, an explicit aws_region_name still wins, and the body carries the bare id AWS expects
* fix(bedrock): keep params AWS refuses natively off the chat completions route
Drop the params each family 400s or 503s on runtime Chat Completions from the native config's supported list (GPT-5.6 penalties, stop, and logprobs, Grok penalties, gpt-oss logit_bias) so drop_params drops them as Converse did, gate legacy functions on GPT-5.6 the same way as tools, and send an Anthropic-style thinking block to Converse, the only route that forwards it
* fix(bedrock): keep schema-less json_object on Converse for the chat completions models
* fix(bedrock): keep every json_object response_format on Converse for the chat completions models
* fix(rust): declare the bedrock runtime chat completions flags on ModelInfo
* fix(bedrock): opt into the native chat completions route through supported_endpoints
* docs(cost-map): describe the bedrock native chat completions capability flags
* revert: docs(cost-map): describe the bedrock native chat completions capability flags
This reverts commit 4101c0ceb2.
cost-map-guard runs main's schema generator under pull_request_target and compares
its output to the PR's committed schema, so a PR that changes the generator's output
cannot pass that required check until the generator change lands on main first. The
descriptions move to a follow-up that lands the generator change ahead of the schema
* test(bedrock): move the native chat completions tests under tests/unit
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): drop reasoning_effort none for grok on the native chat completions route
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): keep converse extension params on the converse route
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): inline http image urls and keep stop on converse for native chat completions
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(bedrock): share the sync remote media inliner
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(image-handling): infer the image mime type when the server sends a generic content type
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(bedrock): stop sending aws_bedrock_project_id as OpenAI-Project on the runtime chat completions route
* feat(bedrock): make native chat completions an opt-in bedrock/chat_completions/ route
Bare Bedrock OpenAI and Grok model ids stay on Converse as on main. The
bedrock/chat_completions/<model> prefix opts a deployment into bedrock-runtime's
/openai/v1/chat/completions, and a request carrying a Converse-only param
still falls back to Converse. The cost map no longer decides the route.
* fix(bedrock): keep chat_completions/-prefixed deployments on the native Responses surface
* fix(bedrock): keep provider response headers on the runtime chat completions route
* feat(bedrock): serve gpt-5.6 and newer on runtime chat completions by default
Unprefixed bedrock/<gpt-5.6+> models whose cost-map row lists /v1/chat/completions
now route to the native OpenAI-compatible endpoint; converse/ pins Converse and
chat_completions/ still opts gpt-oss and Grok in. Guardrails, application inference
profile ARNs, and tools with reasoning keep falling back to Converse per request.
Hoist the remote-media url comprehension into a single-clause helper.
* fix(bedrock): refuse temperature and top_p natively on GPT 5.6 and newer like Converse does
AWS answers temperature and top_p with a 400 on the native Chat Completions endpoint for the GPT 5.6+ models, the same models whose Converse route already dropped both under drop_params via supports_sampling_params: false. The native config now honors that price-map flag, the gpt-6 and gpt-6.1 rows carry it, and the gpt-6 family joins gpt-5 in refusing frequency_penalty, presence_penalty, logprobs, and top_logprobs before the request reaches AWS.
* fix(bedrock): refuse GPT sampling and logprob params natively only while reasoning is on
On bedrock-runtime's native chat completions endpoint, GPT-5.x and GPT-6.x
accept temperature, top_p, frequency_penalty, presence_penalty, logprobs,
and top_logprobs once reasoning_effort is "none", and refuse them with any
other effort or when the effort is unset. The previous commit refused the
sampling params unconditionally from the cost map's supports_sampling_params
flag, which lost the reasoning-off case and never covered the penalties or
logprobs. The refusal now keys on the model being a GPT id and reasoning
being active, raises a 400 UnsupportedParamsError naming the params unless
drop_params drops them, and lets everything through under "none". Grok and
gpt-oss keep their unconditional family refusals.
* refactor(bedrock): keep the Converse route-prefix strip inside the bedrock llms module
* fix(bedrock): forward a non-string reasoning_effort on the native route instead of crashing
A list or dict reasoning_effort hit a frozenset membership test in
without_refused_reasoning_effort and raised TypeError, which the proxy
surfaced as a 500 APIConnectionError with no upstream call. The value is
now left alone unless it is a string Bedrock's native endpoint refuses,
so AWS answers the malformed value with its own 400 like it does for an int
* fix(bedrock): route overlong GPT version digits to Converse and send native chat completions to the runtime endpoint
A model id with more than 4300 version digits raised ValueError in the route check; the digits are now bounded so such ids fall back to Converse. The native chat completions URL now follows Converse's precedence: aws_bedrock_runtime_endpoint (or AWS_BEDROCK_RUNTIME_ENDPOINT) wins over api_base, so a deployment that sets both keeps sending to the same host
* fix(bedrock): route model_id overrides to Converse and never send an empty bearer natively
A deployment whose litellm_params carry model_id (an application inference profile or provisioned throughput ARN) went to the native Chat Completions route with the base model in the URL and model_id left in the body. It now takes Converse like the bedrock/arn:... model form, which encodes the override into the request URL
A blank api_key on a SigV4 deployment became an Authorization header reading Bearer with nothing after it on the native route, since the OpenAI-like header builder writes any non-None key and the signer keeps a non-AWS4 Authorization header. validate_environment now resolves the key through bedrock_bearer_token, so a blank key is signed with SigV4 the way Converse signs it
* test(bedrock): audit the native GPT chat completions route on the integration rig
Adds the /audit cells for the runtime chat completions route: the scripted Bedrock runtime peer, the happy and fallback wire tests, the sad-path and regex worst-case tests, the chaos burst tests, the Messages adapter tests, and the Responses native-route tests. Tests only, no product diff.
* test(bedrock): harden the runtime chat completions audit cells
The chaos peer's shared counter and process now come from the same spawn context, since a fork-context Value handed to a spawn-context process raises on Linux. The peer-kill test waits for the first six answers to reach the client before killing the peer instead of counting accepted requests. The Responses wire tests look the spend row up under both the ciphertext id the caller received and the issued id behind it, matching the chaos file's rule for the pre-encryption row
* fix(bedrock): refuse or drop a non-string reasoning_effort before the native chat completions call
A reasoning_effort sent as an int, a list, or an object on a GPT 5.6+ deployment the native
route serves now answers 400 from litellm before any wire request, naming the type and the
drop_params way out, and is dropped under drop_params so AWS applies its default effort, the
way Converse dropped it on main. The tip since a0cef91f0b forwarded it for AWS to refuse
* test(bedrock): pin router retries off and give the chaos bursts config deployments on an owned proxy
* test(bedrock): wait for the replacement worker before tearing down the sigkill chaos proxy
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mateo <mateo@berri.ai>
* feat(bedrock): drop lookaround regex patterns from tool schemas for Converse models that reject them
* fix(bedrock): rename the lookaround flag to supports_regex_lookaround and keep dropped patternProperties names allowed
The cost-map flag becomes a generic supports_regex_lookaround capability, which the
cost-map schema admits as a supports_* boolean, and the Converse transform now owns
the drop decision instead of the shared tools factory. A patternProperties key dropped
from an object closed by additionalProperties: false leaves its value schema as that
object's additionalProperties, so the names it allowed stay allowed, on the OpenAI
non-Python-regex drop too. tool_with_sanitized_parameters also cleans Anthropic-shape
tools (input_schema).
* fix(router): keep a deployment's supports_regex_lookaround off the shared cost-map entry
A deployment's model_info.supports_regex_lookaround was written to the shared
bedrock/<model> cost-map key, so every sibling deployment of that model id
inherited one deployment's choice. The flag now stays under the deployment's
own id, which is what the Converse lookaround check reads first, and the
shared entry keeps the cost map's value
* test(bedrock): audit the Converse lookaround drop on the wire across endpoints, SDKs, flags and chaos
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* feat(cost_calculator): add cost_per_second for chat per-second pricing
Keep legacy input_cost_per_second and output_cost_per_second as aliases for chat, completion, embedding and responses. When both legacy fields are set, input_cost_per_second wins
Move Bedrock commitment rows to cost_per_second so they bill once
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(cost_calculator): drop legacy per-second fields from chat paths
Keep Azure chat token pricing generic and update inert Voxtral rates and SageMaker examples
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost_calculator): recognize output-only per-second rates
Include output_cost_per_second when checking whether a deployment cost entry has pricing so output-only legacy aliases remain attached to the deployment during cost selection
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(pricing): cover cost_per_second and legacy per-second aliases through the proxy
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(cost_calculator): drop output_cost_per_second as a chat per-second alias
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(cost_calculator): restore output_cost_per_second as a chat per-second fallback
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(cost-map): keep input_cost_per_second on bedrock commitment rows for older clients
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>
Register Sail (providers.json, LlmProviders.SAIL, OpenAI-compatible lists,
ProviderConfigManager) for chat, Responses and /v1/messages, and add its 12
models to both cost maps with asap, balanced and flex price columns.
Sail picks speed and price with metadata.completion_window and rejects
service_tier, so the Sail chat and Responses configs translate the tier:
default and priority to asap, flex to flex, balanced to balanced, auto to no
window. Billing prices the window that was sent. A tier Sail has no window
for, or a window or tier set where billing cannot see it (request metadata,
extra_body), is a 400 unless drop_params is set.
Add balanced to ServiceTier and its _balanced price columns to the model
info types, the Rust catalog and the dashboard schema. A transform_extra_body
hook on the chat and Responses base configs, which returns extra_body
unchanged by default, lets Sail keep the window when a caller also sends
extra_body.metadata. Sail is listed in the Add Model form and model picker.
Co-authored-by: shrey kharbanda <shrey@berri.ai>
* feat(xai): add native xAI batches and files support
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(xai): tighten batch handler typing and avoid Final redeclaration on star import
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* refactor(xai): walk batch result pages iteratively
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(xai): stop paging on empty pagination token and honor litellm.xai_key
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* fix(xai): import NotRequired and TypedDict from typing_extensions for Python 3.10
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* feat(batches): accept image and video endpoints on batch create
* test(xai): lock batch endpoint, auth, and result contracts
The batches test package collided with litellm/batches under pytest prepend, so the All Other Providers shard could not collect the new tests.
* fix(xai): price grok batch usage at xAI's 20 percent batch discount
* fix(xai): map not-found file reads to 404, bill batch reasoning tokens, and add 200k batch tier rates
* refactor(xai): drop routine prose and move tests under tests/unit
* fix(health): hand the resolved provider to list_batches in batch-mode health checks
* test(xai): make tests/unit/llms/xai/batches a package
* fix(xai): walk every page of the files list by pagination_token
---------
Co-authored-by: Mubashir Osmani <mubashir@berri.ai>
Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Co-authored-by: mubashir1osmani <mubashir.osmani777@gmail.com>
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(fal_ai): price nano-banana-2 and nano-banana-pro image generations by resolution
* fix(fal_ai): bill passthrough submits per requested image and register the resolution price keys
---------
Co-authored-by: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
* fix(models): sync openrouter prices from the models API
Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>
* test(models): allow above_32k_tokens cost fields in price map schema test
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(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
DeepSeek charges half the listed rate outside 01:00-04:00 and 06:00-10:00 UTC
Monday to Friday, so every deepseek-flash, deepseek-v4-flash,
deepseek-v4-flash-vision-exp, and deepseek-v4-pro entry now carries an
off_peak_pricing block with those windows and the halved input, output, and
cache-hit rates. The generated cost map schema picks up the block, and the
regression tests pin the peak and off-peak cost of one call at fixed moments.
Google prices Lyria per generated clip, so every Vertex Lyria entry in the
price map now carries a single output_cost_per_image and both the speech
and the passthrough cost paths read that one field. The old
output_cost_per_second and audio_seconds_per_prediction pair assumed a
30 second clip, which does not match the 32.768 second WAV Vertex returns,
and no other model in the map priced audio that way
Drops max_audio_length_hours and max_audio_per_prompt from the price map,
its schema, the generator, and ModelInfo, since nothing reads them, and
drops the audio_mime_type hidden param for the same reason: the response
already carries the resolved content type on its own header
Folds the per-model bundled catalog lookups into one cached parse of the
local cost map, validated with a TypeAdapter over a ReadOnly TypedDict
A gpt-5 model accepts a non-default temperature only while its effective reasoning
effort resolves to "none". litellm had no representation of the effort a model applies
when the request omits reasoning_effort, so it substituted supports_none_reasoning_effort,
which is a different fact. Every model that supports "none" without defaulting to it
therefore had temperature forwarded and rejected upstream, and because the carve-out
returned before the drop_params branch, drop_params: true could not save it.
Declare the fact instead. A new cost-map key, default_reasoning_effort, states the effort
the provider applies when the request omits one, and one shared predicate resolves the
effective effort from it: an explicit reasoning_effort wins, otherwise the declared
default, otherwise the catalogue decides.
That last step matters because the cost map is fetched from the published branch at import
time, so it can be OLDER than the code reading it. On such a map every model looks
undeclared, and reading that as "reasoning is active" would strip temperature from the 39
gpt-5.1/5.2/5.4 entries that accept it, a regression caused by data lag rather than by
anything about the model. So an absent declaration is only meaningful once the catalogue
carries the key at all; a map that predates the feature keeps the answer litellm gave
before it existed, and the conservative answer applies from the moment the data lands.
The top_p/logprobs/top_logprobs gate carried the same assumption spelled differently and
now shares the predicate, as does the Responses API, which reimplemented the rule and is
what the default /v1/messages bridge routes openai models through. Azure normalises its
routing names in one resolver that every capability lookup goes through, which replaces
its bespoke per-lookup rewrite.
Declared on the 37 gpt-5.1/5.2/5.4 entries measured to accept temperature=0 today, so
their behaviour is unchanged. The 23 gpt-5.5/5.6 entries that reject it stay undeclared
and are fixed once the catalogue carries the key.
Resolves LIT-3797
Resolves LIT-5028
Kimi K3 accepts exactly low, high and max, defaults to max, and always thinks.
The map could not say that: medium and high have no supports_*_reasoning_effort
flag because every other reasoning model takes them, so the ten kimi-k3 entries
carried supports_reasoning alone and resolved to unknown. The dashboard then fell
back to a capability-blind level list that deliberately omits max, which is why a
kimi-k3 tier cannot be set to max thinking today.
Add reasoning_effort_levels, an array key in the shape the map already uses for
supported_endpoints and supported_modalities. Where present it is read first and
wins whole; every other entry keeps answering through the per-level flags,
unchanged. It is deliberately a different name from the computed
ModelGroupInfo.supported_reasoning_efforts, which stays derived from a group's
deployments and is never seeded from one deployment's model_info.
The levels are per entry rather than per model, because the deployments differ:
Moonshot, Together, Fireworks and Azure Foundry all forward the level unchanged
and get the model's own low/high/max, while Perplexity documents a six-value
enum it maps down internally and gets that. The /v1/messages degradation chain
consults the same declaration, so the level the map advertises is the level that
path forwards.
Gemini API Maps-grounded prompts were billed as web search and Vertex AI Maps-grounded prompts were not billed at all. Classify grounding metadata per candidate into web search vs Maps requests, carry a distinct google_maps_grounding_requests usage counter through non-streaming and streaming paths, and price it via the new google_maps_grounding_cost_per_query cost map key with per-query and per-prompt defaults keyed off web_search_billing_unit. Fixes#35906
Regenerate model_prices_and_context_window.schema.json and add the flag to
the inline validator schema in test_utils.py so the new cost-map key passes
validate-model-prices-json and the JSON-valid test.
The cache control hook also runs on litellm.responses() input. On a
GPT-5.6 deployment it wrapped a string-content item into a chat-shaped
{"type": "text"} part, which the Responses API rejects, and it never
marked input_text, input_image or input_file parts, so no breakpoint and
no prompt_cache_options reached the provider. Add the Responses part
types to the eligible block set and translate chat-shaped text parts on
non-assistant items to input_text in
ResponsesAPIRequestUtils.merge_prompt_management_input, which both the
async and the sync prompt management sites go through.
The dialect also fired for any GPT-5.6 name that resolved to provider
openai, including deployments pointed at a custom api_base that does not
understand prompt_cache_breakpoint. Decide it once per request from the
provider, the model map and the resolved api_base (request, then
litellm.api_base, then OPENAI_BASE_URL / OPENAI_API_BASE): only
api.openai.com and *.api.openai.com hosts speak the dialect, a top-level
prompt_cache_options opts a custom target in, and litellm_proxy/ targets
never get it. maybe_seed_default_injection_points takes api_base and
stamps the finished decision on the points as _litellm_openai_dialect so
the sync completion() path, whose hook params do not carry api_base,
honors it; maybe_inject_cache_control takes api_base from the
/v1/messages handler.
Eligibility now comes from a supports_prompt_cache_breakpoint model map
flag on the OpenAI gpt-5.6 entries, exposed through
litellm.utils.supports_prompt_cache_breakpoint, with the GPT version rule
kept only for models the map does not know. The OpenAI dialect no longer
reserves a slot for tool_config points, which OpenAI has no cache block
for, and with_prompt_cache_breakpoint plus the chat bridge helper return
a new block instead of mutating their input.
Price ApplyGuardrail usage units recorded by PR #37225 with a new
bedrock/guardrails entry in the model cost map (regional override via
bedrock/{region}/guardrails), add the per-request guardrail_cost to the
standard logging payload's response_cost and CostBreakdown, surface it in
the x-litellm-response-cost header, and bill blocked requests through the
failure hook so key and team budgets see what AWS bills
