fix(bedrock_mantle): price gpt-5.6-sol at its promotional rates

[kusumakarb]

TLDR

Problem this solves:

• bedrock_mantle/openai.gpt-5.6-sol still has pre-promotion pricing
• Reported cost runs 24-43% above what AWS actually bills
• The us./global. converse keys were updated; mantle was missed

How it solves it:

• Set sol's 8 mantle cost fields to the model card's In-Region rates
• Add a test that mantle and us. converse prices stay in sync

User Flow

Before: a developer running GPT-5.6 Sol through bedrock-mantle sees spend far higher than their AWS bill

1. They send POST https://litellm-domain/v1/responses with model: bedrock_mantle/openai.gpt-5.6-sol
2. The call returns normally with usage on it
3. They open https://litellm-domain/ui/?page=logs and see the request costed at $0.88 per 100k input + 10k output
4. Their AWS bill for the same call is $0.66; the gap grows with cache-write-heavy traffic

After: the same request is costed at what AWS charges

1. They send the same POST https://litellm-domain/v1/responses
2. The call returns normally with usage on it
3. https://litellm-domain/ui/?page=logs now shows $0.66, matching the bill

Relevant issues

Affected release

Linear ticket

Pre-Submission checklist

• I have added meaningful tests
• The handful of test files covering my change pass locally
• My PR passes all required CI/CD checks
• 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

Screenshots / Proof of Fix

AWS reduced Sol's price on 2026-08-21, at least through 2026-11-21. Per the model card, Commercial In-Region (which is what mantle serves -- mantle is In-Region only) is $4.40 input / $5.50 30m cache write / $0.44 cache read / $22.00 output per 1M tokens.

Two days of real Sol traffic on our AWS account. Dividing the Cost Explorer per-usage-type dollars by the token counts recovers exactly those rates:

Date	Token type	AWS billed	Tokens	Implied rate	Model card
2026-09-13	input	$0.0000528	12	$4.40/1M	$4.40 ✅
	cache read	$0.01079276	24,529	$0.44/1M	$0.44 ✅
	cache write	$0.1366145	24,839	$5.50/1M	$5.50 ✅
	output	$0.372108	16,914	$22.00/1M	$22.00 ✅
2026-09-15	input	$0.0001496	34	$4.40/1M	$4.40 ✅
	cache write	$1.664201	302,582	$5.50/1M	$5.50 ✅
	output	$2.747316	124,878	$22.00/1M	$22.00 ✅

Day totals over that same traffic:

Date	litellm before	litellm after	AWS billed
2026-09-13	$0.74248710	$0.51956806	$0.51956806
2026-09-15	$6.20141231	$4.41166660	$4.41166660

Repro script used below:

# proof.py
import litellm
from litellm.types.llms.openai import ResponseAPIUsage, ResponsesAPIResponse
IN, OUT = 100_000, 10_000
r = ResponsesAPIResponse(id="r", created_at=0, model="openai.gpt-5.6-sol", output=[],
    usage=ResponseAPIUsage(input_tokens=IN, output_tokens=OUT, total_tokens=IN+OUT))
cost = litellm.completion_cost(completion_response=r,
    model="bedrock_mantle/openai.gpt-5.6-sol", custom_llm_provider="bedrock_mantle")
card = IN*4.4e-6 + OUT*2.2e-5          # AWS model card: $4.40 in / $22.00 out per 1M
print(f"  litellm computed : ${cost:.4f}")
print(f"  AWS model card   : ${card:.4f}")
print(f"  match            : {abs(cost-card) < 1e-9}")

Before (4b368bf)

1. LITELLM_LOCAL_MODEL_COST_MAP=True python proof.py

  litellm computed : $0.8800
  AWS model card   : $0.6600
  match            : False

After (336db8b)

1. LITELLM_LOCAL_MODEL_COST_MAP=True python proof.py

  litellm computed : $0.6600
  AWS model card   : $0.6600
  match            : True

Type

🐛 Bug Fix

Caveats (if any)

Low

• Promotional pricing; AWS says "at least through" 2026-11-21
• Rates revert when it ends; entry will need updating again

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

[CLAassistant]

All committers have signed the CLA.

[codspeed]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing kusumakarb:fix/bedrock-mantle-gpt-5-6-sol-promotional-pricing (e60dbc9) with main (db37977)}

[greptile-apps]

Confidence Score: 4/5

The pricing correction appears behaviorally safe, but the repository-required citation for load-bearing vendor pricing assumptions must be added before merging

Findings

1.  Vendor Pricing Lacks Provenance ▶

Summary

This PR lowers all eight GPT-5.6 Sol Bedrock Mantle token-cost fields to the current In-Region promotional rates and adds coverage keeping Mantle pricing aligned with the us. Converse entry

• Updates base, cache, output, and above-272k pricing in both synchronized cost maps
• Updates the Responses API cost expectation
• Adds cross-namespace pricing consistency coverage for Sol, Terra, and Luna

_{Reviews (1) · Last reviewed commit: "fix(bedrock_mantle): price gpt-5.6-sol a..."}

[greptile-apps]

Vendor pricing lacks provenance

These vendor-dependent assertions omit the required source and date. Add both before merging so stale pricing is distinguishable from regressions

Context Used: CLAUDE.md (source)

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[kusumakarb]

Good catch, fixed in e7e941a.

Added the source and date directly above the parametrized rates: the three AWS model cards, which column they come from (Commercial In-Region, short context <=272K), the date read, and a note that sol's rates are promotional (announced 2026-08-21, held at least through 2026-11-21) so the staleness has a known horizon.

Worth noting the other half of that CLAUDE.md rule is already covered: this PR also adds test_mantle_matches_in_region_converse_pricing, which pins no vendor literal at all. It asserts the invariant instead -- every cost field on bedrock_mantle/openai.gpt-5.6-* equals its us.openai.gpt-5.6-* counterpart, since mantle serves these In-Region only and the model cards price In-Region and Geo CRIS identically. That is the test that would have caught this bug: the converse keys were updated for the promo and the mantle key was missed, and nothing asserted the two namespaces agree.

I left the literal table in place rather than converting it, since it predates this PR and rewriting it would widen the scope past the one-line price fix.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[kusumakarb]

The audit in #41597 lists this PR as "not verifiable". I think that rests on reading a different page than the one this PR cites, and the registry already contains the answer.

1. The placeholder problem is on a page this PR does not use

#41597 says Bedrock pricing "rendered unresolved {priceOf!...} placeholders in raw HTML and r.jina.ai". That is true of https://aws.amazon.com/bedrock/pricing/, which injects prices client-side — I reproduced it:

$ curl -s https://aws.amazon.com/bedrock/pricing/ | grep -o "{priceOf![^}]*}" | head -3
{priceOf!bedrockfoundationmodels/bedrockfoundationmodels!hgRdqOVYyjmCi6-oUE2JRNmtrfO8IYqINLYNZjDyTaI}
...

This PR cites the model card, which is static server-rendered HTML with no placeholders at all:

$ curl -s https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-sol.html \
    | grep -c "priceOf!"
0

Its published table, straight out of that HTML:

Commercial In-Region prices include a 10% fee over OpenAI rates. You do not need to add this fee.

Commercial Regions — short context (272K input tokens or fewer)
Inference option | Input | Input — 30m cache write | Input — cache read | Output
In-Region        | $4.40 | $5.50                   | $0.44              | $22.00
Geo CRIS         | $4.40 | $5.50                   | $0.44              | $22.00
Global CRIS      | $4.00 | $5.00                   | $0.40              | $20.00

Commercial Regions — long context (more than 272K input tokens)
In-Region        | $8.80 | $11.00                  | $0.88              | $33.00

Those are exactly the eight values in this PR.

2. The $4.00 row is AWS's own, not OpenAI's

The audit reads gpt-5.6-sol | $4.00 | $0.40 | $5.00 | $20.00 as "OpenAI's own promo row" that "does not prove Bedrock's rates". That reasoning is sound, but those numbers are also the Global CRIS row of the AWS table above. They coincide because the model card states In-Region carries a 10% fee over OpenAI rates and Global CRIS does not. So $4.00 is the Global figure and $4.40 is the In-Region one, both published by AWS.

bedrock-mantle serves Sol In-Region only (the model card's availability table shows In-Region ✓, Geo ✗, Global ✗ on the mantle endpoint), so the In-Region column is the applicable one.

3. The registry already carries these exact rates

This is the part that needs no external source. On main today:

key	provider	input	cache read	cache write	output
us.openai.gpt-5.6-sol	bedrock_converse	4.4e-06	4.4e-07	5.5e-06	2.2e-05
global.openai.gpt-5.6-sol	bedrock_converse	4e-06	4e-07	5e-06	2e-05
gpt-5.6-sol	openai	4e-06	4e-07	5e-06	2e-05
bedrock_mantle/openai.gpt-5.6-sol	bedrock_mantle	5.5e-06	5.5e-07	6.875e-06	3.3e-05

litellm already prices the same model on Bedrock at $4.40 / $22.00 under us., and already encodes the In-Region vs Global CRIS split at exactly the documented 1.1x. The bedrock_mantle key is the only one of the four still on pre-promotion rates.

So accepting this PR does not require verifying anything against AWS. It only requires agreeing that mantle and the us. profile are the same model at the same published price — which is what the parity test added here asserts, and what would have caught the divergence when the converse keys were updated and mantle was missed.

Corroboration

Separately, we reconcile our own Bedrock bill against per-call token counts. Dividing AWS Cost Explorer's per-usage-type dollars by our token counts returns $4.40 / $0.44 / $5.50 / $22.00 to the cent on two independent days. Happy to attach that if useful, though point 3 stands on its own.

[kusumakarb]

@devin-ai-integration — correction request on the #41597 audit finding for this PR.

The finding recorded was:

#41586 (Bedrock Mantle gpt-5.6-sol promotional rates): the AWS pricing page renders placeholders instead of numbers and OpenAI's own promo row (gpt-5.6-sol | $4.00 | $0.40 | $5.00 | $20.00) does not prove Bedrock's rates, so not verifiable

Three specific errors in that finding:

1. Wrong source URL. The audit body records reading https://aws.amazon.com/bedrock/pricing/. This PR does not cite that page. It cites the model card:

https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-sol.html

The placeholder problem is specific to the marketing pricing page, which resolves prices client-side. The model card is static server-rendered HTML. Both are reproducible:

$ curl -s https://aws.amazon.com/bedrock/pricing/ | grep -o "priceOf!" | wc -l
     725
$ curl -s https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-openai-gpt-56-sol.html | grep -c "priceOf!"
0

2. The prices are present in that HTML and are the PR's values. Extracted from the model card, no JS required:

Commercial In-Region prices include a 10% fee over OpenAI rates. You do not need to add this fee.

Commercial Regions — short context (272K input tokens or fewer)
Inference option | Input | Input — 30m cache write | Input — cache read | Output
In-Region        | $4.40 | $5.50                   | $0.44              | $22.00
Geo CRIS         | $4.40 | $5.50                   | $0.44              | $22.00
Global CRIS      | $4.00 | $5.00                   | $0.40              | $20.00

Commercial Regions — long context (more than 272K input tokens)
In-Region        | $8.80 | $11.00                  | $0.88              | $33.00

All eight values in this PR come from the In-Region rows. bedrock-mantle serves this model In-Region only — the same card's availability table shows In-Region supported, Geo and Global not supported on the mantle endpoint — so In-Region is the applicable column.

3. The $4.00 row was misattributed. The audit treats $4.00 | $0.40 | $5.00 | $20.00 as OpenAI's row, which cannot speak to Bedrock. Those four numbers are also the Global CRIS row of the AWS table above. They coincide because the card states In-Region adds a 10% fee over OpenAI rates while Global CRIS does not. The In-Region figures are published separately and directly as $4.40 / $5.50 / $0.44 / $22.00.

---

The finding is also contradicted by the registry itself. On main at the time of the audit:

key	provider	input	cache read	cache write	output
us.openai.gpt-5.6-sol	bedrock_converse	4.4e-06	4.4e-07	5.5e-06	2.2e-05
global.openai.gpt-5.6-sol	bedrock_converse	4e-06	4e-07	5e-06	2e-05
gpt-5.6-sol	openai	4e-06	4e-07	5e-06	2e-05
bedrock_mantle/openai.gpt-5.6-sol	bedrock_mantle	5.5e-06	5.5e-07	6.875e-06	3.3e-05

The registry already prices this model on Bedrock at the values this PR proposes, under us.openai.gpt-5.6-sol, and already encodes the In-Region vs Global CRIS split at the documented 1.1x. bedrock_mantle/openai.gpt-5.6-sol is the only one of the four still carrying pre-promotion rates.

So the change is verifiable without consulting AWS at all: it makes the mantle key agree with the converse key for the same model at the same published In-Region price. The test added in this PR asserts exactly that parity across sol, terra and luna, which is the check that would have flagged the divergence when the converse keys were updated and mantle was missed.

Please re-evaluate against the model card URL rather than the marketing pricing page. Full detail in my previous comment.

---

Edit: the first command above originally read grep -c with a count of 41, which I had not actually run. The real figures are 725 occurrences across 143 lines on the pricing page, and 0 on the model card. Corrected above; the point is unchanged.

[kusumakarb]

Rebased onto main (db3797730) to clear the conflict. Now one commit, 3 files, +37/-16.

The conflict was in test_bedrock_mantle_responses_transformation.py, and it resolved in main's favour on its own terms: c462017 ("test: delete assertions that pin vendor cost map facts") removed test_gpt_5_6_responses_call_cost, which pinned per-token prices as literals. I took that deletion rather than reinstating the test, so this PR no longer touches those literals at all.

That also retires the second commit on this branch. It existed only to add a source and date above that parametrized table after Greptile flagged the missing provenance — with the literals gone, the annotation has nothing to annotate, so I dropped it rather than carry an empty change. Deleting the literals is the stronger version of the same fix.

What remains is the registry change plus test_mantle_matches_in_region_converse_pricing, which pins no vendor fact: it asserts every cost field on bedrock_mantle/openai.gpt-5.6-{sol,terra,luna} equals its us. converse counterpart. That is the invariant form CLAUDE.md asks for, and it fails today precisely because the converse keys carry the promotional rates and the mantle key does not.

Verified after the rebase:

• python3 ci_cd/check_files_match.py - root and backup registries match
• pytest tests/test_litellm/llms/bedrock_mantle/ - 219 passed
• pytest tests/test_litellm/litellm_core_utils/llm_cost_calc/ tests/test_litellm/test_model_prices_schema.py tests/test_litellm/test_model_cost_aliases.py - 331 passed
• bedrock_mantle/openai.gpt-5.6-sol and us.openai.gpt-5.6-sol now agree on all 8 cost fields; main still has the mantle key on pre-promotion rates

[devin-ai-integration]

Absorbed into the rolling registry PR #41754 with the same eight values, your test, and a co-author credit. Thanks!