About ModelPriceWatch

Some provider sign-up links on this site are affiliate links. If you sign up through them we may earn a commission at no extra cost to you. This never affects which models we list or how we rank them — pricing data is always shown cheapest-first based on blended cost.

The badge next to a model means its price is traced to that provider's official pricing page — the source link shown on every model page, so you can check the number yourself in one click — and is re-verified on a schedule, not entered once and forgotten. Two checks run automatically: daily, every price we can map is compared against an independent public reference (the open-source LiteLLM price table) and any divergence is flagged for a human to reconcile against the official page; and weekly, the highest-value models are re-read directly from the provider's own pricing page, capturing a dated receipt. The badge is a provenance signal, not an endorsement — it tells you the number came from, and is reconciled against, the provider's own page, not a third-party guess. If we ever can't tie a price to an official source, we drop the model rather than show an unsourced number.

• 218 Current — listed on the provider's own page as of the date shown on each model page.
• 1 Preview — listed, but the provider labels it preview or experimental; pricing can change without notice.
• 13 Deprecated — the provider has announced end-of-life. Still purchasable, still in our tables, badged so you don't start new work on one.
• 38 Retired — no longer served on the endpoint we price. Removed from the tables, kept at their own URLs for historical reference, and clearly marked.

We re-check these against provider deprecation pages and correct them in the open. In July 2026 that audit found seven models we had been listing as current which providers had already shut down — the oldest more than a year earlier. Those are now marked Retired with their end dates.

Shown as 1633.3 from August 28, 2026 to August 28, 2026. Withdrawn on August 28, 2026; we publish no value for it now.

Right number, wrong league. LMArena runs several separate leaderboards — text, web development, vision and others — and a model's Elo on one is not comparable with an Elo on another, because each is scored only against the models in that board's own pool. Every LMArena Elo we publish is a text-leaderboard rating. This model has no text rating: it appears on the web-development board only, and that board's 1633.3 was published in the text column. Shown next to text ratings it read as this model outscoring the highest-rated models we track, which the underlying data does not say. We have withdrawn the figure rather than substitute one, because no text rating for this model exists yet; it will appear here once the leaderboard publishes one.

Where it appeared: this model's own page, the LMArena Elo board on the rankings page, comparison rows against this model.

Knock-on effects:

• This model's percentile average across benchmarks, as at the correction: 99.4 → withdrawn.
• Percentile average across benchmarks, other models, as at the correction: 66 rows shifted, by a median of 0.1 and at most 1.

No published price was affected. Unlike our other benchmark columns, this Elo IS part of the percentile average we rank on, so removing it moved other models' percentile figures — all of them upward, by the amounts shown, because the withdrawn rating had been sitting near the top of the population. No price, and no other benchmark score, was affected. This model now shows no benchmark scores at all, which is the honest position until a source publishes one.

Shown as 34 from July 18, 2026 to August 8, 2026. Corrected to 42.2 on August 8, 2026.

Since superseded: we publish 25.8 as of September 8, 2026. Artificial Analysis has since replaced the index itself. This is not a further correction of ours: the August 8 fix was about which of Artificial Analysis's entries we read, and that join is unchanged and still correct — the same generally-available Hy3 entry, still marked measured rather than estimated. What changed is the scale. On September 7, 2026 Artificial Analysis published Intelligence Index v4.3, which drops one component benchmark, adds a harder one in its place and upgrades a third to a new version. Re-basing a composite moves every model scored on it, and it did: the new index rolled across their directory over September 7 and 8, and of the models we track that carry it, every one moved downward — 18 on the first day and 67 on the second, none upward. Our figure went 42.2 to 32.4 to 25.8 across those two days for that reason. A score on v4.3 is not comparable to a score on v4.2, and neither reading says the model got worse. We record it here so this correction cannot be read against a current figure it was never about.

Our row's id and name say "preview", but the SKU it prices is Tencent's generally-available Hy3 release. The index was joined to Artificial Analysis's separate, deprecated April-2026 preview entry, and the score on that entry is one Artificial Analysis itself marks as estimated rather than measured. We published it as an independently measured score.

Where it appeared: this model's own page, the Artificial Analysis Intelligence Index board on the rankings page, comparison rows against this model.

Knock-on effects:

• This model's percentile average across benchmarks, as at the correction: 33.1 → 51.1.
• Percentile average across benchmarks, other models, as at the correction: 69 rows shifted, by a median of 0.1 and at most 0.6.

No published price was affected. The index is a display-only composite and is deliberately excluded from the percentile average we rank on, so the ranking tables moved only by the amounts shown above.