feat(models): add GPT-5.6 (sol/terra/luna) pricing and metadata

[devin-ai-integration]

Relevant issues

Linear ticket

Pre-Submission checklist

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

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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)

Screenshots / Proof of Fix

Captured at commit 1f6e8640ce

A live completion against the real OpenAI API wasn't possible at capture time: GPT-5.6 is rolling out globally over roughly 24 hours, so GET https://api.openai.com/v1/models/gpt-5.6 (and -sol) still return 404 for this account even with a valid key, and a chat request 404s upstream. The end-user-visible behavior this PR actually changes is the pricing and capability metadata the gateway resolves and serves, so the proof is a real litellm proxy on localhost:4000 booted with the updated cost map, queried over HTTP with no mocks. A live completion was re-run once the account gained 5.6 access; see the update at the end of this section

Boot:

OPENAI_API_KEY=... LITELLM_LOCAL_MODEL_COST_MAP=True \
  uv run python litellm/proxy/proxy_cli.py --config gpt56_proof_config.yaml --port 4000

curl -s http://localhost:4000/model/info (summarized) returns the new entries with the pricing and metadata from this PR:

gpt-5.6 -> openai/gpt-5.6
  input/out per tok: 5e-06 / 3e-05
  cache read/write : 5e-07 / 6.25e-06
  >272k in/out     : 1e-05 / 4.5e-05
  max_in/max_out   : 1050000 / 128000
  reasoning/vision/web: True True True
gpt-5.6-sol -> openai/gpt-5.6-sol
  input/out per tok: 5e-06 / 3e-05
  cache read/write : 5e-07 / 6.25e-06
  >272k in/out     : 1e-05 / 4.5e-05
  max_in/max_out   : 1050000 / 128000
  reasoning/vision/web: True True True
gpt-5.6-terra -> openai/gpt-5.6-terra
  input/out per tok: 2.5e-06 / 1.5e-05
  cache read/write : 2.5e-07 / 3.125e-06
  >272k in/out     : 5e-06 / 2.25e-05
  max_in/max_out   : 1050000 / 128000
  reasoning/vision/web: True True True
gpt-5.6-luna -> openai/gpt-5.6-luna
  input/out per tok: 1e-06 / 6e-06
  cache read/write : 1e-07 / 1.25e-06
  >272k in/out     : 2e-06 / 9e-06
  max_in/max_out   : 1050000 / 128000
  reasoning/vision/web: True True True

The cache-write rate is 1.25x the uncached input rate for every tier, the long-context (>272K) rate is 2x input and 1.5x output, and the alias gpt-5.6 matches gpt-5.6-sol, all as intended

Update, captured at commit 594c61bc86: the account now has GPT-5.6 access, so here is the fully live end-to-end proof, hitting the real OpenAI API through the proxy with no mocks and costing real $. Same boot command as above; one completion per model, printing the message content, token usage, and the proxy's x-litellm-response-cost header:

for m in gpt-5.6 gpt-5.6-sol gpt-5.6-terra gpt-5.6-luna; do
  curl -sD headers.txt http://localhost:4000/v1/chat/completions \
    -H 'Authorization: Bearer sk-1234' -H 'Content-Type: application/json' \
    -d "{\"model\": \"$m\", \"messages\": [{\"role\": \"user\", \"content\": \"Reply with exactly: hello from $m\"}]}" \
    | jq -c '{model, content: .choices[0].message.content, prompt_tokens: .usage.prompt_tokens, completion_tokens: .usage.completion_tokens}'
  grep -i '^x-litellm-response-cost:' headers.txt
done

Output:

{"model":"gpt-5.6","content":"hello from gpt-5.6","prompt_tokens":18,"completion_tokens":11}
x-litellm-response-cost: 0.00042
{"model":"gpt-5.6-sol","content":"hello from gpt-5.6-sol","prompt_tokens":19,"completion_tokens":12}
x-litellm-response-cost: 0.000455
{"model":"gpt-5.6-terra","content":"hello from gpt-5.6-terra","prompt_tokens":20,"completion_tokens":13}
x-litellm-response-cost: 0.000245
{"model":"gpt-5.6-luna","content":"hello from gpt-5.6-luna","prompt_tokens":20,"completion_tokens":13}
x-litellm-response-cost: 9.8e-05

Every cost matches this PR's rates exactly: luna is 20 x 1e-06 + 13 x 6e-06 = 9.8e-05, terra is 20 x 2.5e-06 + 13 x 1.5e-05 = 0.000245, sol is 19 x 5e-06 + 12 x 3e-05 = 0.000455, and the bare gpt-5.6 alias prices at the sol rates as intended (18 x 5e-06 + 11 x 3e-05 = 0.00042)

Type

🆕 New Feature

Changes

Day 0 support for OpenAI's GPT-5.6 family in the model cost map. Adds gpt-5.6, gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna to both model_prices_and_context_window.json and the bundled litellm/model_prices_and_context_window_backup.json, so cost tracking, context-window enforcement, and capability flags work with no code changes. The bare gpt-5.6 alias mirrors the flagship gpt-5.6-sol tier

No transformation code was needed. OpenAIGPT5Config.is_model_gpt_5_model already routes anything matching gpt-5.* through the GPT-5 reasoning path, and is_model_gpt_5_4_plus_model parses the version numerically so 5.6 lands on the 5.4-or-newer side that auto-switches to the /v1/responses bridge when tools are combined with reasoning_effort

Pricing follows OpenAI's published rates (per 1M tokens, standard tier):

• sol: $5 input, $30 output, $0.50 cached input, $6.25 cache write
• terra: $2.50 input, $15 output, $0.25 cached input, $3.125 cache write
• luna: $1 input, $6 output, $0.10 cached input, $1.25 cache write

Each entry carries the full tier matrix the cost calculator understands: the standard, Flex, Batch, and Priority service tiers, plus the above-272K-token long-context variants for input, output, and cache read. Cache writes are new for this family and are billed at 1.25x the uncached input rate, encoded as cache_creation_input_token_cost (with _above_272k_tokens, _flex, and _priority variants). The family shares a 1,050,000-token input window, 128,000-token max output, reasoning support, web search, vision, structured outputs, and the 10% regional-processing uplift that applies to models released on or after March 5, 2026

Reasoning-effort flags mirror GPT-5.5 (OpenAI's guide states GPT-5.5 prompting guidance carries over to GPT-5.6): none and xhigh supported, minimal not supported. The newer max reasoning effort and pro reasoning mode surfaced in the GPT-5.6 guide are follow-ups; they need transformation work and are out of scope here

Commit 594c61bc86 cleans three floating point representation artifacts in the luna cache-read rates flagged by Greptile, so the raw JSON now carries the intended 2e-07 and 5e-08 instead of values like 2.0000000000000002e-07

Scope is OpenAI only. azure_ai/gpt-5.6 variants can follow once Azure exposes the models

Tests

• tests/test_litellm/test_gpt_5_6_model_metadata.py pins pricing, capability flags, endpoints, modalities, and provider routing for all four names, and asserts the main and backup maps stay in sync
• test_generic_cost_per_token_gpt56 in test_llm_cost_calc_utils.py checks standard, long-context, and cache-write pricing flow through generic_cost_per_token
• the is_model_gpt_5_model regression list and a new is_model_gpt_5_4_plus_model class both cover the gpt-5.6 family

Link to Devin session: https://app.devin.ai/sessions/93a883e6bc334fd4892564f4a2b14797
Requested by: @mateo-berri

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

This PR adds day-0 pricing and capability metadata for the GPT-5.6 model family (gpt-5.6, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna) to both model_prices_and_context_window.json and its bundled backup, enabling cost tracking and context-window enforcement with no production code changes.

• Pricing entries include the full tier matrix (standard, flex, batch, priority, above-272K long-context) and the new cache-write cost field (cache_creation_input_token_cost), with values that are mathematically consistent across all tiers and models.
• Tests in test_gpt_5_6_model_metadata.py, test_llm_cost_calc_utils.py, and test_is_model_gpt_5_model.py cover pricing, capability flags, provider routing, and the is_model_gpt_5_4_plus_model classification; all tests use local files or mocked data with no network calls.
• test_utils.py's schema validator is extended to accept the three new cache_creation keys (_above_272k_tokens, _flex, _priority) introduced by this family.

Confidence Score: 5/5

Safe to merge — changes are purely additive JSON data and matching tests with no production logic modifications.

The change is entirely confined to JSON cost-map entries and new/extended tests. All pricing multipliers are mathematically consistent, both cost-map files stay in sync, and no production code paths are modified.

No files require special attention.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Adds four new GPT-5.6 model entries with complete pricing tiers; values are numerically consistent with the stated pricing rules (cache-write = 1.25x input, long-context input = 2x, long-context output = 1.5x)
litellm/model_prices_and_context_window_backup.json	Mirror of main cost map; sync verified by the new test_gpt_5_6_backup_matches_main test
tests/test_litellm/test_gpt_5_6_model_metadata.py	New test file verifying pricing, capability flags, endpoints, modalities, and provider routing for all four GPT-5.6 names; no network calls, reads local JSON files
tests/test_litellm/litellm_core_utils/llm_cost_calc/test_llm_cost_calc_utils.py	Adds parametrized cost-calculation tests for gpt-5.6 family; uses LITELLM_LOCAL_MODEL_COST_MAP so no network calls are made
tests/test_litellm/llms/openai/test_is_model_gpt_5_model.py	Extends GPT5_MODELS list with gpt-5.6 variants and adds a new TestOpenAIGPT5ConfigIsModelGpt54PlusModel class to exercise the version-gating logic; all pure unit tests
tests/test_litellm/test_utils.py	Schema validator extended to accept the three new cache_creation cost keys; additive change that does not weaken existing validation

_{Reviews (2): Last reviewed commit: "fix: floating point entry errors" | Re-trigger Greptile}

[emerzon]

Can you please also add the Azure entries?

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[mateo-berri]

@greptileai

[mateo-berri]

@emerzon thanks for the reminder. We're adding this in a follow-up

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_gpt_5_6_rollout (594c61b) with litellm_internal_staging (131aa05)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (0099c6b) during the generation of this report, so 131aa05 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩