fix(model_prices): add bedrock_mantle gpt-5.5/5.4 272K tiers, align sol with AWS invoice

[mateo-berri]

TLDR

Problem this solves:

• Bedrock gpt-5.5 and gpt-5.4 prompts over 272K tokens logged at the flat rate, half the invoice
• gpt-5.6-sol logged at promo rates the AWS invoice does not bill, 20% under

How it solves it:

• Adds the *_above_272k_tokens input, output, and cache-read rates to gpt-5.5 and gpt-5.4
• Moves sol base and tier rates to the invoiced $5.50 / $33.00 / $0.55 / $6.88 per million
• Replaces the test that pinned the flat behaviour with one that pins the invoiced numbers

User Flow

Before: a platform team reconciling gateway spend against its AWS bill finds Bedrock GPT-5.5 long prompts logged at half of what AWS charges, and Sol at 20% under

1. Their app sends POST https://litellm-domain/v1/responses with "model": "mantle-55" (a bedrock_mantle/openai.gpt-5.5 deployment) and a 300K-token input
2. It gets 200 with usage.input_tokens: 301016, and the response headers say x-litellm-response-cost-input: 1.655588, $5.50 per million for the whole prompt
3. The same call to mantle-54 (bedrock_mantle/openai.gpt-5.4) comes back with x-litellm-response-cost-input: 0.827794, $2.75 per million
4. A short prompt to mantle-sol (bedrock_mantle/openai.gpt-5.6-sol) with 16 input and 5 output tokens returns x-litellm-response-cost-input: 7.04e-05 and x-litellm-response-cost-output: 0.00011, $4.40 and $22.00 per million
5. https://litellm-domain/ui/?page=logs shows the requests at that spend, while the AWS Cost Explorer lines for the same day bill the 300K prompt under input_tokens_long_ctx_standard at $11.00 per million (gpt-5.5) and $5.50 (gpt-5.4), and Sol at $5.50 input and $33.00 output, so the gateway under-reports the gpt-5.5 prompt by $1.66 and every Sol call by 20%

After: the same calls come back priced at what the AWS invoice charges

1. Their app sends the same POST https://litellm-domain/v1/responses with "model": "mantle-55" and the 300K-token input
2. It gets 200 with usage.input_tokens: 301015, and now x-litellm-response-cost-input: 3.311165, $11.00 per million for the whole prompt
3. The same call to mantle-54 comes back with x-litellm-response-cost-input: 1.6555825, $5.50 per million
4. The short prompt to mantle-sol returns x-litellm-response-cost-input: 8.8e-05 and x-litellm-response-cost-output: 0.000165, $5.50 and $33.00 per million
5. https://litellm-domain/ui/?page=logs shows the requests at those amounts, matching the Cost Explorer lines

Relevant issues

Linear ticket

Resolves LIT-6356

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• The handful of test files covering my change pass locally, e.g. uv run pytest tests/test_litellm/<your_test_file>.py -v. Leave the suites (make test-unit-*, make test-unit) to CI: it finishes in ~15 minutes where a laptop takes an hour or more
• My PR passes all required CI/CD checks (e.g., lint, schema.d.ts sync check, etc.)
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Shared setup: each leg is its own DB-less proxy booted with 2 uvicorn workers (python litellm/proxy/proxy_cli.py --config lit6356_config.yaml --port <port> --num_workers 2) from a worktree at that leg's commit, with LITELLM_LOCAL_MODEL_COST_MAP=True and AWS_BEARER_TOKEN_BEDROCK for the Bedrock CI account, on this config. The streaming case that closes each side ran on a separate DB-backed proxy (real Postgres, same config, 2 workers): base on port 43917, head on port 41873

model_list:
  - model_name: mantle-55
    litellm_params:
      model: bedrock_mantle/openai.gpt-5.5
      aws_region_name: us-east-1
  - model_name: mantle-54
    litellm_params:
      model: bedrock_mantle/openai.gpt-5.4
      aws_region_name: us-east-1
  - model_name: mantle-sol
    litellm_params:
      model: bedrock_mantle/openai.gpt-5.6-sol
      aws_region_name: us-east-1
general_settings:
  master_key: sk-1234

The 300K bodies carry the same plain-text filler prompt on every endpoint (it asks for the single word "pong"), prefixed with a per-leg nonce so Bedrock's prompt cache stays cold: every call below reports cached 0. responses_*.json puts it in the Responses input string, chat_54.json in a chat messages list, and messages_54.json in an Anthropic messages list with max_tokens: 64. Each step runs curl -s -D hdr.txt -o out.json -w "http %{http_code}\n" <route> -H 'Authorization: Bearer sk-1234' -H 'Content-Type: application/json' --data-binary @<body> and then prints grep -i 'x-litellm-response-cost' hdr.txt and the usage object from out.json

What AWS bills the same account, from Cost Explorer for 2026-08-24 to 08-26: gpt-5.5 input_tokens_long_ctx_standard $11.00 per million and output_tokens_long_ctx_standard $49.50, gpt-5.4 $5.50 and $24.75, sol input_tokens_standard $5.50, output $33.00, cache read $0.55, cache write 30m $6.88

Before (98c5233)

gpt-5.5, 300K prompt, POST /v1/responses

1. curl ... http://localhost:42685/v1/responses --data-binary @lit6356_qa_before_responses_55.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.5
x-litellm-response-cost-input: 1.655588
x-litellm-response-cost-output: 0.000165
x-litellm-response-cost: 1.655753
usage: input 301016, cached 0, output 5, text 'pong'

3. 301016 x $5.50 per million; AWS bills 301016 x $11.00 per million = $3.311176 input, twice what was logged

gpt-5.4, 300K prompt, POST /v1/responses

1. curl ... http://localhost:42685/v1/responses --data-binary @lit6356_qa_before_responses_54.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.4
x-litellm-response-cost-input: 0.827794
x-litellm-response-cost-output: 8.25e-05
x-litellm-response-cost: 0.8278765
usage: input 301016, cached 0, output 5, text 'pong'

3. 301016 x $2.75 per million; AWS bills $5.50 per million = $1.655588 input

gpt-5.4, 300K prompt, POST /v1/chat/completions

1. curl ... http://localhost:42685/v1/chat/completions --data-binary @lit6356_qa_before_chat_54.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.4
x-litellm-response-cost-input: 0.827794
x-litellm-response-cost-output: 8.25e-05
x-litellm-response-cost: 0.8278765
usage: input 301016, cached 0, output 5, text 'pong'

gpt-5.4, 300K prompt, POST /v1/messages

1. curl ... http://localhost:42685/v1/messages --data-binary @lit6356_qa_before_messages_54.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.4
x-litellm-response-cost-input: 0.827794
x-litellm-response-cost-output: 8.25e-05
x-litellm-response-cost: 0.8278765
usage: input 301016, cached None, output 5, text 'pong'

gpt-5.6-sol, short prompt, POST /v1/responses

1. curl ... http://localhost:42685/v1/responses -d '{"model":"mantle-sol","input":"Reply with exactly the word pong and nothing else."}'
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.6-sol
x-litellm-response-cost-input: 7.04e-05
x-litellm-response-cost-output: 0.00010999999999999999
x-litellm-response-cost: 0.0001804
usage: input 16, cached 0, output 5, text 'pong'

3. $4.40 input and $22.00 output per million; AWS bills $5.50 and $33.00, so $8.8e-05 and $1.65e-04

gpt-5.4, 300K prompt, POST /v1/chat/completions with stream: true, then GET /spend/logs

1. curl -N ... http://localhost:43917/v1/chat/completions --data-binary @lit6356_risk_base_stream_chat_54.json (the chat_54 body plus "stream": true and "stream_options": {"include_usage": true})
2. Output:

http 200
usage chunk: prompt_tokens 301025, completion_tokens 5, cached 0, text 'pong'

3. curl -s "http://localhost:43917/spend/logs?request_id=<the row's resp_... id, from http://localhost:43917/ui/?page=logs>" -H 'Authorization: Bearer sk-1234'
4. Output:

{'model': 'bedrock_mantle/openai.gpt-5.4', 'spend': 0.8279012499999999, 'prompt_tokens': 301025, 'completion_tokens': 5, 'total_tokens': 301030, 'call_type': 'responses'}

5. 301025 x $2.75 per million plus 5 x $16.50 per million lands in the spend row, the flat rate

After (837bcba)

gpt-5.5, 300K prompt, POST /v1/responses

1. curl ... http://localhost:42414/v1/responses --data-binary @lit6356_qa_after_responses_55.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.5
x-litellm-response-cost-input: 3.311165
x-litellm-response-cost-output: 0.0002475
x-litellm-response-cost: 3.3114125
usage: input 301015, cached 0, output 5, text 'pong'

3. 301015 x $11.00 per million input and 5 x $49.50 per million output, the invoiced long-context rates

gpt-5.4, 300K prompt, POST /v1/responses

1. curl ... http://localhost:42414/v1/responses --data-binary @lit6356_qa_after_responses_54.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.4
x-litellm-response-cost-input: 1.6555825
x-litellm-response-cost-output: 0.00012375
x-litellm-response-cost: 1.65570625
usage: input 301015, cached 0, output 5, text 'pong'

3. 301015 x $5.50 per million input and 5 x $24.75 per million output

gpt-5.4, 300K prompt, POST /v1/chat/completions

1. curl ... http://localhost:42414/v1/chat/completions --data-binary @lit6356_qa_after_chat_54.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.4
x-litellm-response-cost-input: 1.6555825
x-litellm-response-cost-output: 0.00012375
x-litellm-response-cost: 1.65570625
usage: input 301015, cached 0, output 5, text 'pong'

gpt-5.4, 300K prompt, POST /v1/messages

1. curl ... http://localhost:42414/v1/messages --data-binary @lit6356_qa_after_messages_54.json
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.4
x-litellm-response-cost-input: 1.6555825
x-litellm-response-cost-output: 0.00012375
x-litellm-response-cost: 1.65570625
usage: input 301015, cached None, output 5, text 'pong'

gpt-5.6-sol, short prompt, POST /v1/responses

1. curl ... http://localhost:42414/v1/responses -d '{"model":"mantle-sol","input":"Reply with exactly the word pong and nothing else."}'
2. Output:

http 200
x-litellm-model-name: bedrock_mantle/openai.gpt-5.6-sol
x-litellm-response-cost-input: 8.8e-05
x-litellm-response-cost-output: 0.000165
x-litellm-response-cost: 0.00025299999999999997
usage: input 16, cached 0, output 5, text 'pong'

3. $5.50 input and $33.00 output per million, matching the invoice

gpt-5.4, 300K prompt, POST /v1/chat/completions with stream: true, then GET /spend/logs

1. curl -N ... http://localhost:41873/v1/chat/completions --data-binary @lit6356_risk_head_stream_chat_54.json (the chat_54 body plus "stream": true and "stream_options": {"include_usage": true})
2. Output:

http 200
usage chunk: prompt_tokens 301025, completion_tokens 5, cached 0, text 'pong'

3. curl -s "http://localhost:41873/spend/logs?request_id=<the row's resp_... id, from http://localhost:41873/ui/?page=logs>" -H 'Authorization: Bearer sk-1234'
4. Output:

{'model': 'bedrock_mantle/openai.gpt-5.4', 'spend': 1.65576125, 'prompt_tokens': 301025, 'completion_tokens': 5, 'total_tokens': 301030, 'call_type': 'responses'}

5. 301025 x $5.50 per million plus 5 x $24.75 per million lands in the spend row, the invoiced long-context rates

Observations from the run:

• Whole completion bills the 1.5x output tier past 272K, as AWS does
• /v1/messages usage omits cache_read_input_tokens; this PR leaves it alone
• Cache-read tiers not exercised live: every call reported cached 0
• Streaming spend row carries the tier on head
• Spend row request_id is the resp_ id on both legs

Type

🐛 Bug Fix

Caveats (if any)

Medium

• AWS announced Sol promo pricing ($4.40 / $22.00) on 2026-08-21, through at least 2026-11-21
  • Cost Explorer for 08-24, 08-25, and 08-26 still bills $5.50 / $33.00 / $0.55 / $6.88, with no credit lines
  • This PR follows the invoice; fix(pricing): add promo pricing for GPT-5.6 Sol on Bedrock #38127 follows the announcement and goes the other way on sol
  • If the meter drops to the promo rate, sol needs a follow-up back to the 4.4e-06 rates
• us.openai.gpt-5.6-sol and global.openai.gpt-5.6-sol keep the promo rates, no invoice data for them

Low

• The 2x cache-read tier for gpt-5.5 and gpt-5.4 mirrors OpenAI's own gpt-5.5 tier shape
  • Cost Explorer shows no cache-read line for either model yet, even after cached replays
  • Luna's invoiced long-context cache-read line is 2x its standard one, same family shape
• The 272K threshold mirrors sol and OpenAI direct; only 300K-plus prompts were sent
• AWS's gpt-5.5 model card still lists flat pricing and a 272K window; the invoice disagrees
• make check gates none of these files; ruff on the two test files reports only pre-existing findings

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

• 837bcba passes /live-pr-risk

[greptile-apps]

Greptile Summary

The PR aligns Bedrock Mantle GPT-5 pricing with observed AWS invoice rates.

• Adds above-272K input, cache-read, and output tiers for GPT-5.5 and GPT-5.4.
• Updates GPT-5.6 Sol base, long-context, cache-read, cache-write, and output rates.
• Keeps the root and bundled pricing maps synchronized and expands cost-calculation regression coverage.

Confidence Score: 5/5

The PR appears safe to merge with the pricing maps synchronized and the revised base and long-context behavior covered by tests.

The metadata keys match the generic threshold-pricing contract, both runtime map copies agree, and no concrete changed-code failure remains.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Updates the canonical Bedrock Mantle GPT-5.4, GPT-5.5, and GPT-5.6 Sol pricing entries consistently with the intended invoice rates.
litellm/model_prices_and_context_window_backup.json	Mirrors the canonical pricing changes in the bundled runtime fallback map without divergence.
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py	Replaces the obsolete flat-rate expectation with base-boundary and long-context cost assertions and adds Sol cache-write rate coverage.
tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py	Updates Bedrock Mantle model-info and end-to-end response cost expectations for the revised Sol rates.

_{Reviews (1): Last reviewed commit: "fix(model_prices): add bedrock_mantle gp..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 837bcba. Configure here.}

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_mantle_gpt5_cost_tiers (837bcba) with litellm_internal_staging (98c5233)}