The LLM market moves. Since the last verification pass roughly a month ago, two providers shipped new flagships, one cut prices fleet-wide, another retired its previous-generation IDs and started silently redirecting calls, and a fifth doubled the price of three older models. v0.27.0 is the catalog refresh that catches up.
New IDs are now recognized by dippin lint. New prices flow through dippin cost. Stale IDs come out so authors get a DIP108 nudge instead of routing requests at a model that doesn’t quite exist anymore.
OpenAI shipped the GPT-5.5 line plus a slate of -pro variants:
| Model | Input | Output |
|---|---|---|
gpt-5.5 | $5.00 | $30.00 |
gpt-5.5-pro | $30.00 | $180.00 |
gpt-5.4-pro | $30.00 | $180.00 |
gpt-5.2-pro | $21.00 | $168.00 |
gpt-5 | $1.25 | $10.00 |
gpt-5-pro | $15.00 | $120.00 |
gpt-5-mini | $0.25 | $2.00 |
gpt-5-nano | $0.05 | $0.40 |
xAI shipped Grok 4.3 — grok-4.3 at $1.25 / $2.50. DeepSeek rolled out V4 as deepseek-v4-flash ($0.14 / $0.28) and deepseek-v4-pro ($1.74 / $3.48 list; in a 75% launch discount through May 31). Mistral added mistral-medium-3-5-2604 ($1.50 / $7.50, new flagship-class), mistral-medium-3-1-2508 ($0.40 / $2.00), and the Ministral 3 generation (ministral-3-3b-2512, ministral-3-8b-2512, ministral-3-14b-2512). Google promoted gemini-3.1-flash-lite to GA. Cohere has new Command-A variants in the docs (reasoning, vision, translate) but no public per-token pricing yet — they’re not in the catalog until prices surface.
The single most consequential change: OpenAI doubled the price of every legacy GPT model that still has a price. gpt-5.2, gpt-5.1, and gpt-4.1-mini all went up 2x on both input and output. The newer -mini/-nano tiers and the o-series held steady. If you had workflows pinned to those three models, your dippin cost estimates were off by 2x until you bumped to this version.
xAI went the other direction. The entire grok-4.20-*-0309 family came down from $2/$6 to $1.25/$2.50, matching grok-4.3’s tier. Three models, fleet-wide cut.
DeepSeek’s V4-Flash is half the price of the legacy aliases on input ($0.14 vs $0.28) and a third less on output ($0.28 vs $0.42). The deepseek-chat and deepseek-reasoner aliases now resolve to V4-Flash and bill at V4-Flash rates — dippin cost reflects that.
Anthropic repriced Haiku 3.5 to $0.80 / $4.00. The model was retired on the first-party API in February but remains available via Bedrock and Vertex AI; the new rate is what Anthropic now publishes on the pricing page.
Pin this somewhere.
Already retired (removed from the catalog):
| Date | Model |
|---|---|
| 2026-04-30 | mistral-small-3.2 |
| 2026-03-09 | gemini-3-pro-preview |
| 2026-02-27 | pixtral-large |
Retired but still functionally callable (kept in the catalog, priced at the rate they’re actually billed):
| Date | Model |
|---|---|
| 2026-05-15 | grok-4-1-fast-reasoning, grok-4-1-fast-non-reasoning — xAI silently redirects to grok-4.3 and bills at grok-4.3 rates ($1.25/$2.50). Requests still execute, so DIP108 would be misleading. Migrate when convenient. |
Coming up (still in catalog with deprecation comments):
| Date | Model |
|---|---|
| 2026-06-01 | gemini-2.0-flash |
| 2026-06-15 | claude-sonnet-4-0, claude-opus-4-0 |
| 2026-07-24 | deepseek-chat, deepseek-reasoner aliases (sunset to V4-Flash) |
| 2026-10-23 | gpt-4o, gpt-4.1-nano, o3-mini, o4-mini |
The 2026-06-01 Gemini date is the urgent one — that’s two weeks out. Workflows still pinned to gemini-2.0-flash should plan to migrate to gemini-2.5-flash-lite or gemini-3.1-flash-lite.
A note on the two categories. Hard-retired models (provider returns an error) come out of the catalog so DIP108 surfaces — calling them is a real bug. Soft-retired models that the provider silently redirects (like xAI’s grok-4-1-fast-* family) stay in, priced at the rate the redirect target actually bills. Treating them like typos would be wrong twice over: the request still executes, and cost analysis needs the real billed price, not $0. A future, more specific “deprecated alias — prefer X” diagnostic can replace the current behavior of staying silent on these.
A few values we couldn’t verify cleanly from canonical sources this pass, held at prior values pending re-check:
nemo and mistral-small-2603 — Mistral’s pricing tab is JS-rendered and didn’t yield to our doc fetcher. Third-party scrapers disagree on the prices. Values held at $0.02/$0.04 and $0.10/$0.30 respectively until manual verification.command-a-03-2025 and command-r7b-12-2024 — Cohere removed per-token pricing for these from the public page. Held at $2.50/$10.00 and $0.0375/$0.15.dippin cost models the base tier only.All flagged with inline code comments for the next pass.
One quieter change worth noting. Cohere’s documentation now lists command-r-08-2024, command-r-plus-08-2024, and command-r7b-12-2024 as the canonical IDs; the bare command-r/command-r-plus/command-r7b aliases resolve to versions that were deprecated 2025-09-15. The catalog now accepts both the dated and bare forms (so existing .dip files keep validating), but the dated IDs are recommended. Bare aliases carry deprecation comments.
If you need a model we don’t catalog yet — provider-specific betas, fine-tuned variants, a hosted gateway with custom IDs — register them at startup:
validator.RegisterExtraModels("openai:my-custom-model;anthropic:claude-internal-2026")
Unknown provider:model combos produce a DIP108 warning, not a hard error, so workflows still run — but adding the ID makes the warning go away and gets you accurate cost estimation.
Sources for every provider are pinned in the validator/lint_model.go and cost/pricing.go source-file comments. We verify against canonical provider docs only — never third-party aggregators, never LLM training data. (Training data is always stale by definition; the rule is in the project’s CLAUDE.md.) Verification cadence is roughly monthly. If you spot a model on a provider’s official docs that we don’t recognize, file an issue with the canonical URL.
Full changelog: CHANGELOG.md.