LLM API Price Report: July 2026

This is the first in a monthly series. Each report covers one calendar month of our own daily price snapshots — what changed, on which dates, and by how much. Nothing here is forecast or commentary on models we don’t track; if a month is quiet, the report will say so and be short.

A note on timing: this first report is published in September, compiled from snapshots we recorded daily throughout July. The data was collected on the day; only the write-up is retrospective. Future reports will follow their month directly.

July 2026 was not quiet.

Coverage: 31 of 31 days, no gaps, from 2026-07-01 to 2026-07-31, with the 2026-06-30 snapshot used as the baseline so changes landing on the 1st are caught.

The month in one table

Four out of five models did not move at all. Among the ones that did, increases outnumbered decreases by nearly four to one.

The headline: OpenAI’s new mid-tier lost 90% of its price in three weeks

OpenAI’s GPT-5.6 family appeared in our data on 2026-07-10 at these output prices:

Then, in the last four days of the month:

Luna ended July at one tenth of its launch price, 21 days after it appeared. Sol, the flagship of the three, did not move at all in July — it stayed at $30 all month.

This is the single most useful thing a daily record catches, and it is worth being precise about why. A model that launches at one price and settles at another within a month leaves no trace of the first number anywhere. There is no changelog. Anyone who evaluated Luna in mid-July, wrote down $6, and built a cost model around it was working from a number that was wrong by 10× before the month ended — and would have had no way to discover that except by checking again.

(These launches are also why our main movers table above misses them: a model that wasn’t in the June 30 baseline can’t have a “net monthly change.” We track mid-month launches separately for exactly this reason.)

Anthropic shipped three models and repriced nothing

Anthropic added three entries in July:

And changed the price of zero existing models: 0 of 14 Anthropic models we tracked at both ends of July moved a cent.

That combination — ship new, never reprice old — is a distinct strategy, and it showed up again in August. It has a practical consequence: on Anthropic, your bill changes when you migrate, not when they decide something. The counterpart is that you get no passive price relief either; the model you’re on today will cost the same next month.

The Opus 5 “Fast” variant is worth noting on its own. It carries exactly double the price of standard Opus 5 on both input and output ($10/$50 vs $5/$25) for the same 1M context. Speed is being sold as a 2× multiplier, priced as cleanly as batch pricing is discounted.

July 16th: the day the market repriced

Change events are not spread evenly. One date dominated:

Twenty-eight models across at least four different labs changed output price on July 16th. Some of the month’s largest single moves land on that date: Qwen3 Coder 480B’s cut, Gemma 3 27B’s rise, Llama 3.1 8B’s rise.

We can’t see why from pricing data alone, and we won’t pretend to. What we can say is that a weekly or monthly price check would have shown you the aftermath of July 16th as a set of unexplained differences, with no way to tell one lab’s decision from a coordinated market shift.

Biggest moves among models present all month

Down:

Only ten models ended July cheaper than they started it. That is the entire “prices are falling” story for the month.

Up:

Read these honestly: they start from very small absolute numbers, and a move from $0.10 to $3.00 per million output tokens is a model going from implausibly cheap to merely cheap. This is normalization of introductory pricing, not inflation.

But the practical effect on a budget is identical to inflation. A team that benchmarked Qwen3 235B Thinking at $0.10 in late June and sized its spend on that number was paying 30× more by July 31st, with no migration and no notification.

Qwen repriced half its catalogue; OpenAI touched 4%

The split is stark and it is not about price level — it’s about policy. The labs offering the cheapest models are the ones repricing them constantly.

Note the asymmetry with the headline story: OpenAI touched only 2 of 54 tracked models, yet still produced the month’s largest price collapse — because the collapse happened on models that launched mid-month and therefore aren’t in this table at all.

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

GLM 5.2 changed price 55 times in 31 days and finished 1.2% from where it began. Kimi K2.7 Code changed 16 times and ended exactly flat. For models like these, a single price lookup is close to meaningless — you need a distribution, not a number.

The market median barely moved

Comparing only the 276 models present on both June 30 and July 31, the median output price went from $1.50 to $1.56 — up 4.0%.

⚠️ One caution on that figure: it is a like-for-like median over that specific set of models. It is not comparable to the same statistic in another month, because the model population changes. We publish it as a within-month reading only.

What didn’t happen

Batch pricing was completely stable in July: zero changes in the batch-to-standard ratio across every batch/standard pair we track. Every batch variant that was priced at 50% of standard when it appeared was still at 50% on July 31st.

(That changed sharply in August — six ratio breaks — which the next report covers.)

One thing we can’t tell you

On 2026-07-29, 28 :batch listings appeared in our data at once, including Anthropic’s entire tracked line-up. We are not going to tell you 28 batch endpoints launched that day, because we can’t know that from price data. The far more likely explanation is that our data source began exposing batch pricing it hadn’t surfaced before.

This is a real limitation of what we do: we record what our source publishes, and a change in coverage looks identical to a change in the market. Where we can’t distinguish the two, we say so rather than picking the more interesting story. It’s the same reason we exclude -latest aliases from our counts — an alias repointing to a new model looks exactly like a price change, and isn’t one.

Takeaways

• 80% of the market didn’t move. Stability is the default, even in an active month.
• Increases beat cuts 38 to 10 among models tracked all month.
• The biggest move was on a brand-new model: GPT-5.6 Luna, −90% in 21 days from launch. Launch prices are the least reliable numbers in this market.
• Anthropic shipped three models and repriced none. Different labs, genuinely different policies.
• July 16th repriced 28 models across four labs. Weekly checks would have missed the event and seen only the residue.
• Nothing here is recoverable after the fact. Every number above exists only because it was written down on the day.

Based on computetrail’s daily snapshots from 2026-07-01 to 2026-07-31 (31 consecutive days, no gaps), with 2026-06-30 as the baseline. Counts exclude -latest aliases and :batch variants unless stated; see our methodology. Large percentage moves on low-priced models usually reflect repricing or normalization rather than a market-wide trend.