LLM API price changes, September 2026: who cut, who doubled

The short version

• Some output prices fell sharply. DeepSeek V4 Pro 0423 went from $2.05 to $0.463 per 1M output tokens (-77%), the largest decrease among the models listed here.
• Others doubled. GPT-5.6 Sol Pro went from $10.00 to $20.00 (+100%), and GLM 5.3 Flash from $0.25 to $0.50 (+100%).
• The consequence: a cost estimate based on an old price snapshot can become inaccurate in either direction. Recheck the price of every model you run, every month.

Unless stated otherwise, all prices below are output prices per 1M tokens, comparing the Obolith snapshots of 2026-08-31 and 2026-10-01.

The cuts

Prices as of 2026-10-01, compared with 2026-08-31:

DeepSeek models account for most of this list.

The ordering between DeepSeek models also changed. On 2026-08-31, DeepSeek V3.1 had the lower output price ($1.65 vs $2.05 for DeepSeek V4 Pro 0423). On 2026-10-01, the order was reversed: DeepSeek V4 Pro 0423 at $0.463, DeepSeek V3.1 at $0.95. If output price was your reason for staying on DeepSeek V3.1, that reason no longer holds on output price alone. Input prices, quality and latency still need their own check.

At unchanged output-token volume, these lower rates reduce the output portion of your cost estimate.

The increases

Same window, same unit:

The increases hit both high and low price points. GLM 5.3 Flash and Granite 4.2 8B were already among the lower output prices in this comparison, so their changes look small in absolute dollars. In percentage terms they are not: at unchanged output-token volume, a +100% output price means +100% on the output portion of your cost for that model.

Where the new prices now sit, on output price only:

• GPT-5.6 Sol Pro at $20.00 matches Claude Opus 5.5 ($20.00). GPT-6.1 Sol lists output at $10.00, which is what GPT-5.6 Sol Pro cost on 2026-08-31.
• GLM 5.3 Flash at $0.50 matches GPT-6 Luna ($0.50) and is above Llama 3.3 70B Instruct ($0.32).
• Qwen3.5 397B A17B at $3.50 is above DeepSeek V4 Pro 0423 ($0.463).

These comparisons cover token prices only; they do not establish equivalent quality, latency, or workload cost.

The snapshots show which tracked output prices changed between 2026-08-31 and 2026-10-01. They do not establish why those prices changed, and we will not guess.

Models first tracked during the comparison window

These models were not in the 2026-08-31 snapshot, so there is no earlier price to compare against. First tracked means first recorded by Obolith, not necessarily newly released. Prices as of 2026-10-01:

Claude Sonnet 5.5 and GPT-6.1 Sol have identical tracked prices ($2.00 input, $10.00 output). Price does not separate them; context (1M vs 1.1M) and your own evals will.

Both were first recorded on the last day of the window, so each has a single data point here. Treat those prices as a starting point, not a trend.

What to do with this

A price change only matters if it reaches your cost model. Depending on what you run:

• DeepSeek V4 Pro 0423, Llama 3.3 70B Instruct, DeepSeek V4 Flash 0423 or DeepSeek V3.1: the tracked output price went down. If your forecast still uses the 2026-08-31 output rate, update that line with the current billed rate and your expected output-token volume.
• GPT-5.6 Sol Pro, GLM 5.3 Flash, Granite 4.2 8B or Qwen3.5 397B A17B: the tracked output price went up. If your forecast still uses the 2026-08-31 output rate, update it the same way. Before accepting the new rate, look at the alternatives in the tables above and run them through your own evals.
• Choosing a model now: the models first tracked in this window add options, but only token prices are compared here.

A monthly routine:

1. Open LLM API pricing and find every model you run in production.
2. Compare today's input and output prices with the ones in your cost model.
3. If a price moved, confirm the rate your provider actually bills, then recompute with your real input and output token volumes.
4. If you are weighing a rented GPU against the API, rerun that arithmetic with the self-host vs API calculator; the reasoning is in Is it cheaper to self-host, or just use an API?.

If you are stuck on a model whose output price went up, check the levers that do not require a migration, listed in Six levers to cut your LLM API bill. Be precise about what each one touches:

• Batch API: for non-urgent jobs, check whether your provider's batch lane offers a lower output rate for that model.
• Prompt caching: it can reduce the input portion of your bill when requests share a long opening block (system prompt, document, tool schema). It does not reduce the output-token rate. Whether input savings offset an output-price increase in your total bill depends on your input and output volumes and the applicable caching rates.

Method

• Daily collection. Obolith records LLM API prices every day.
• Normalization. Prices are expressed per 1M tokens, input and output separately.
• Comparison window. This article compares the 2026-08-31 snapshot with the 2026-10-01 snapshot. The change tables report output prices; input-price changes are not covered here. Intermediate days are not analyzed in this article, so it does not state on which day each change happened.
• First-tracked models are those first recorded by Obolith during the window, with the date of first collection.

These are the prices tracked by Obolith. Caching and batch pricing can affect your effective cost. Confirm the rate your provider actually bills before updating your forecast.

Takeaway

Update the output-price line in your forecast, not the entire bill by the same percentage. Confirm your provider's billed rate first.