LLM API Price Report: August 2026

Second in our monthly series, covering 31 of 31 days with no gaps, 2026-08-01 to 2026-08-31, with the 2026-07-31 snapshot as baseline. (July’s report is here.)

August was busier than July in events but calmer in direction: more price changes spread over fewer models, and they cancelled out.

The month in one table

Two things stand out against July. The catalogue grew for the first time (34 in, 17 out, net +17). And the near four-to-one bias toward increases disappeared — August’s movers split almost evenly, with 19 more models moving and ending exactly where they began.

More events on fewer models means the activity concentrated: a smaller group of models repricing more often.

August 18th: OpenAI moved six models, in both directions, on one day

This is the event of the month, and a snapshot comparison would render it as noise.

(Output $/1M. The :batch variants of Sol and Sol Pro tracked the cut the same day, $15 → $7.50.)

On one date, the same lab halved its flagship tier and doubled its two cheaper tiers. Four days later Sol went again, $15 → $10 on 08-22, finishing August at one third of where it started.

Set against July, this reads as a correction of a correction. Luna launched at $6 on 07-10, fell to $0.60 by 07-31 (−90%), and then doubled back to $1.20 on 08-18. Terra went $15 → $6 in July, then back to $12 in August. Meanwhile Sol, untouched all July at $30, took the month’s largest cut.

The whole GPT-5.6 line was repriced twice in seven weeks, in opposite directions, and settled at:

Anyone who benchmarked this family in July was working with a price list that no longer describes any of the three models.

Three providers ended their batch discount on the same day

July had zero changes to the batch-to-standard price ratio. August had six, and half of them landed on one date.

Note on the last row: DeepSeek V4 Flash 0731’s batch listing entered our data already priced above standard, so that row is the ratio settling rather than a discount being withdrawn. We include it for completeness, not as evidence of a provider decision.

On 2026-08-19, three unrelated providers raised their batch output price to exactly match standard. In each case the standard price didn’t move — only the batch rate did. Anyone who had recorded “batch saves 50%” on these models in July was carrying a stale assumption into September.

NVIDIA’s went further: after losing its discount on the 19th, Nemotron 3 Ultra’s standard price fell to $2.20 on 08-28 while the batch rate stayed at $3.60 — leaving the batch endpoint 64% more expensive than the regular one.

The Gemini 3.7 Flash entry is the mirror image. Its standard price doubled on 08-29 ($1.875 → $3.75) but the batch price didn’t follow until 09-03, so for five days batch ran at a quarter of standard rather than half — an unadvertised discount that existed purely in the gap between two repricings.

We pulled this thread further in a separate piece on how reliable the half-price rule actually is: Batch APIs: Is It Really Half Price?

A coverage caveat

Two more waves of :batch listings appeared in our data during August — 32 on 08-07 and 17 on 08-29 — following 28 on 2026-07-29. As in July’s report, we read these as our data source widening what it exposes, not as dozens of simultaneous product launches. Price data alone can’t separate a change in coverage from a change in the market, so we don’t guess.

The six ratio changes above are a different matter: those are pairs we had already been recording on both sides, where the relationship between two prices we were watching demonstrably changed.

Biggest moves

Down (models present all month):

Qwen3 14B is the mirror of July: it rose 279% in July ($0.24 → $0.91) and gave all of it back on a single day in August, returning to exactly $0.24. Two months of daily records show a round trip that either month alone would have reported as a dramatic one-way move.

Up:

Two launches that repriced within days

DeepSeek V4 Pro 0813 launched at $0.87, hit $3.564 within eight days — a 4× move — and settled at $1.98, more than double its launch price. Gemini 3.7 Flash simply doubled 15 days after appearing.

Combined with July’s Luna story, a clear pattern is now visible across two months of records: the price a model launches at is the least durable number in this market. Four of the six mid-month launches that repriced in August did so within nine days.

The market median finished exactly flat

Across the 291 models present on both 2026-07-31 and 2026-08-31, the median output price was $1.75 at the start and $1.75 at the end — 0.0% change.

⚠️ As in July’s report: this is a like-for-like median over that specific set, and is not comparable across months because the population changes. (July’s reading, $1.50 → $1.56, covered a different 276 models.) Treat each as a within-month statement only.

A flat median alongside 377 change events is the point worth keeping: a lot happened, and in aggregate none of it moved the middle of the market.

Provider behaviour held

Two months in, the ordering is stable enough to plan around. DeepSeek repriced over half its catalogue in both months. Anthropic has now gone 62 consecutive days without changing a single published price — while shipping three new models in July. OpenAI’s rate quadrupled, entirely because of August 18th.

Qwen is the notable mover: from 53% of its catalogue in July down to 32% in August, after a July that included several large normalizations.

Some models still don’t have a price, they have a range

Kimi K2.6 changed price 24 times in August and finished at exactly its starting number — as it nearly did in July (12 changes, +25%). GLM 5.2 has now logged 93 price changes across our two reported months.

Takeaways

• Activity concentrated: fewer models moved than in July (53 vs 60) but they moved more often (377 events vs 350).
• Direction cancelled out. 16 up, 18 down, 19 moved-and-returned; the market median finished flat at $1.75.
• August 18th: one lab, six models, two opposite directions, one day.
• Batch discounts are not permanent. Three vanished on 08-19; one endpoint is now dearer than standard.
• Launch prices don’t hold. Both August launches that moved did so within nine days, one by 4×.
• Anthropic: 62 days, zero repricing. The provider you pick determines whether your unit costs are a fixed input or a moving one.

Based on computetrail’s daily snapshots from 2026-08-01 to 2026-08-31 (31 consecutive days, no gaps), with 2026-07-31 as the baseline. Counts exclude -latest aliases and :batch variants unless stated; see our methodology. Monthly median figures are like-for-like within each month and are not comparable between months.