OpenAI cut GPT-5.6 Luna's API price by 80% on July 30, 2026, to $0.20 per million input tokens and $1.20 per million output tokens. DeepSeek is moving the other way. On July 31, the day after the cut, DeepSeek attached a footnote to its own pricing table announcing a peak/off-peak policy at twice the regular rate. Between August 5 and August 6 it pulled that footnote and replaced it with a broader one: it plans to raise API prices "in the near future, with a significant increase expected."
Put those two facts side by side and the standard mental model breaks. The cheap tier is not permanently Chinese. On today's published rates, Luna costs 1.43x DeepSeek V4-Flash on input tokens and 4.29x on output. Not the order of magnitude that gets quoted in every cost-comparison thread. And on OpenAI's Batch tier, Luna's input token is already cheaper than DeepSeek's.
If your unit economics assume a 10x saving from switching to DeepSeek, the number is stale by roughly a week.
What actually changed, and on which date
On July 30, OpenAI published "Advancing the price-performance frontier with GPT-5.6." The operative sentence: "Starting July 30, API pricing is $2 per million input tokens and $12 per million output tokens for Terra, and $0.20 per million input tokens and $1.20 per million output tokens for Luna. Sol pricing remains unchanged." The archived rate card from July 29, the day before the cut, lists Luna at $1.00 input and $6.00 output and Terra at $2.50 and $15.00. That confirms the drops as 80% and 20%, and Sol as unchanged at $5.00 and $30.00.
On July 31, DeepSeek shipped DeepSeek-V4-Flash-0731 and moved the V4-Flash API into public beta. The changelog entry covers agent benchmarks, the public-beta status, native Responses API support and Codex adaptation. It says nothing about price.
The price signal never went in the changelog. It went into footnote 2 of the Models and Pricing page, with no date and no headline, in two different versions inside a week.
The capture from July 31 at 07:38 UTC is the first to carry a pricing footnote at all. It read: "The DeepSeek API service will soon adopt a peak/off-peak pricing policy. During peak hours, prices will be 2x the regular prices, applicable to all billing items. The effective date will be subject to the official announcement." The peak window it named was 9:00 to 12:00 and 14:00 to 18:00 Beijing time, daily. The 02:30 UTC capture that same morning does not have it, and neither does any earlier capture we pulled, back to July 6. A concrete doubling landed within about a day of OpenAI's cut.
That text was still there in the capture from August 5 at 09:56 UTC. On the live page on August 6 it is gone, replaced in the same footnote slot by something vaguer: "We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly. The specific pricing plan will be subject to official notice." The word peak no longer appears anywhere on the page.
So the sequence is a specific 2x surcharge posted the day after OpenAI cut, then withdrawn inside roughly 24 hours and replaced by an open-ended warning. DeepSeek has not explained the swap, and nothing we retrieved makes either footnote a response to OpenAI. For a cost model the useful part is this: the rate card carries a stated upward direction with no number and no date, and the last number the vendor put in writing was a doubling.
The 10x discount is gone. Here is the real multiplier
Published rates as of August 6, 2026, per million tokens. OpenAI figures from the developer pricing page, DeepSeek from its Models and Pricing page. The ratio is Luna divided by V4-Flash, so above 1.0 means OpenAI costs more.
| Tier (per 1M tokens) | GPT-5.6 Luna | DeepSeek V4-Flash | Ratio |
|---|---|---|---|
| Standard input, short context | $0.20 | $0.14 | 1.43x |
| Standard output, short context | $1.20 | $0.28 | 4.29x |
| Cached input / cache hit | $0.02 | $0.0028 | 7.14x |
| Cache write | $0.25 | not billed separately | n/a |
| Standard input, long context | $0.40 | $0.14 | 2.86x |
| Standard output, long context | $1.80 | $0.28 | 6.43x |
| Batch input † | $0.10 | no batch tier | 0.71x |
| Batch output † | $0.60 | no batch tier | 2.14x |
| Fast mode input † | $0.40 | no equivalent | 2.86x |
| Fast mode output † | $2.40 | no equivalent | 8.57x |
† DeepSeek publishes no batch tier and no fast tier. Those four ratios divide OpenAI's tier rate by DeepSeek's standard rate, so they are cross-tier, not like-for-like.
The single most quoted comparison, standard output tokens, is 4.29x. Every other row is a different number. There is no longer one multiplier between these products, but one per tier, ranging from 0.71x to 8.57x depending on how you call the API.
On the Batch tier, DeepSeek is no longer the cheap input
OpenAI's Batch tier prices Luna at $0.10 per million input tokens. DeepSeek V4-Flash costs $0.14 on a cache miss, and publishes no batch discount at all. For latency-tolerant, input-dominated work, think classification, extraction, document tagging, OpenAI is now 29% cheaper per input token than the model everyone treats as the budget floor.
Output tokens still favour DeepSeek at 2.14x on that tier. But "cheapest API" is no longer a single answer, and for a large class of high-volume work it is no longer DeepSeek. Check this row first, because it is the one that quietly reverses. This is a rate-card comparison, not a benchmark, and quality per dollar is a separate question.
Where the gap widens again: long context and cache hits
The compression is not uniform. Two patterns push it back the other way.
OpenAI tiers its pricing by context length. Luna's long-context input rate is double the short-context rate and its output rate is 1.5x, at $0.40 and $1.80. The same 2x input, 1.5x output pattern holds for Terra and Sol, so it is OpenAI's structure rather than a quirk of the cheap tier. DeepSeek V4-Flash charges one flat rate across its full 1M-token context window. So the moment your prompts cross OpenAI's short-context threshold, the output gap goes from 4.29x to 6.43x.
Where that threshold sits takes reading. The pricing table prints "(<272K context length)" against the legacy gpt-5.5 and gpt-5.4 rows, but does not restate that figure for the GPT-5.6 family. If 272K is the boundary there too, the 100K-input example below stays on short-context rates and most workloads never cross the line. Check your own bills rather than assume the label carries across.
Caching is the sharper one. DeepSeek prices a cache hit at $0.0028, a 50x discount off its own cache-miss rate. OpenAI prices Luna's cached input at $0.02, a 10x discount. On cache hits specifically, DeepSeek is 7.14x cheaper, wider than the headline gap. OpenAI also charges $0.25 per million tokens to write to cache, which DeepSeek does not appear to bill separately.
That matters because agent loops are cache-heavy by construction, replaying a long system prompt and a growing scratchpad every turn. The more agentic your workload, the more of your spend lands where DeepSeek's advantage is largest. Anyone comparing DeepSeek and ChatGPT on the headline rate is measuring the tier that flatters OpenAI most.
The compression happened at the bottom tier only
This is the constraint on the whole thesis. Nothing converged at the middle or the top. GPT-5.6 Terra runs $2.00 input and $12.00 output. DeepSeek V4-Pro runs $0.435 input and $0.87 output. That is 4.6x on input and 13.8x on output. GPT-5.6 Sol, unchanged at $5.00 and $30.00, sits 11.5x and 34.5x above V4-Pro.
So the honest version is narrow. OpenAI's cheapest model closed most of the distance to DeepSeek's cheapest. Above that line the order-of-magnitude gap is intact. If you are choosing a mid-tier or frontier model, nothing this week changes your arithmetic, and the same applies when weighing Claude or Qwen against either vendor.
Why OpenAI says it could cut, and why you cannot check it
OpenAI's stated cause is unusually specific. Within a human-led process, GPT-5.6 Sol "autonomously rewrote and optimized production kernels, designed and ran hundreds of experiments to improve token generation." The company says "the kernel work helped reduce the end-to-end cost of serving the model by 20%, while its experiments increased token-generation efficiency by more than 15%."
Take that at face value and it describes a compounding loop: the model makes itself cheaper to serve, which funds the next cut. Take it sceptically and note that a 20% serving-cost reduction does not obviously fund an 80% price cut. The rest has to come from somewhere, and margin compression during a competitive window is the ordinary explanation. Neither reading is verifiable from outside.
DeepSeek's side is equally unverifiable. Its headline agent scores for the new Flash, Terminal Bench 2.1 at 82.7 and DeepSWE at 54.4, were produced using "DeepSeek Harness minimal mode (to be released soon)" at max effort level. The harness is not released, so nobody outside DeepSeek can reproduce those numbers today. Treat them as vendor claims, not measurements.
What to actually do with your cost model
Three moves, in order of how much money they move.
Recompute per task, not per token. At 10K input and 1K output on a cache miss, Luna costs $0.0032 per call and Flash $0.00168, a 1.9x gap. At 20K input and 20K output, an output-heavy agent turn, Luna is $0.028 and Flash is $0.0084, a 3.3x gap. At 100K input with 90% cache hits plus 10K output, Luna is $0.0158 and Flash is $0.0045, a 3.5x gap. None of those is 10x, and none is 4.3x either. Your number depends on your shape.
Price the switching cost honestly. DeepSeek serves an OpenAI-compatible endpoint and an Anthropic-format endpoint at api.deepseek.com/anthropic, and V4-Flash is adapted for Codex, so migration is closer to a config change than a rewrite. That cuts both ways: low switching cost also means low lock-in, which is why neither vendor can hold a price premium for long.
Do not model a beta rate as a floor. V4-Flash is still in public beta, V4-Pro's official release "will follow soon," and DeepSeek has told you in writing that a significant increase is coming with no date attached. The previous version of that footnote named a specific multiple. Build the sensitivity in now. If you need cheap inference at predictable cost, our roundup of free alternatives to expensive AI tools covers the fixed-cost options, and our DeepSeek vs ChatGPT comparison covers the capability side that pricing alone will not tell you.
One caution on sourcing. DeepSeek rewrote its own price footnote inside a single day this week, so trackers describing a 2x peak-hours surcharge were quoting the vendor accurately when they wrote it, and are now describing text the vendor has removed. Put a date and timestamp on every DeepSeek price you work from, including these.
FAQ
How much did OpenAI cut GPT-5.6 Luna's price? By 80%, effective July 30, 2026. The new rate is $0.20 per million input tokens and $1.20 per million output tokens on the standard short-context tier. GPT-5.6 Terra was cut 20% in the same announcement, to $2.00 and $12.00. GPT-5.6 Sol was not cut and remains at $5.00 and $30.00. OpenAI said the lower Luna and Terra rates also flow through to subscription metering, so the same usage now consumes fewer credits in Codex and ChatGPT Work.
Is DeepSeek really raising its API prices? DeepSeek has said so twice in one week, in two forms. On July 31, 2026 its pricing page gained a footnote announcing a peak/off-peak policy at 2x the regular rate during Beijing business hours, effective date to be announced. By August 6 that footnote was gone, replaced in the same slot by a broader statement that it plans to raise overall API pricing in the near future, with a significant increase expected and the plan subject to official notice. The direction is stated. The magnitude and timing are not.
What is the real cost gap between GPT-5.6 Luna and DeepSeek V4-Flash? There is no single number any more. On standard short-context rates, Luna is 1.43x on input and 4.29x on output. On cache hits the gap widens to 7.14x in DeepSeek's favour, and on long context to 6.43x on output. On OpenAI's Batch tier, Luna's input token is 29% cheaper than DeepSeek's. Compute it for your own input-to-output ratio and cache-hit rate rather than quoting a headline multiple.
Is DeepSeek still 10x cheaper than OpenAI? Not against Luna. The archived OpenAI rate card from July 29, 2026 lists Luna at $1.00 input and $6.00 output, which put it 7.1x DeepSeek V4-Flash on input and 21.4x on output. After the cut those are 1.43x and 4.29x. The 10x framing still roughly holds higher up: GPT-5.6 Terra is 13.8x DeepSeek V4-Pro on output tokens, and Sol is 34.5x. The cheap tiers converged. The mid-tier and frontier gaps did not move.
Why did OpenAI cut the price? OpenAI attributes it to efficiency gains, stating that GPT-5.6 Sol autonomously rewrote and optimized production kernels, cutting end-to-end serving cost by 20% and raising token-generation efficiency by more than 15%. That is the company's own account and cannot be verified externally. It also does not fully explain an 80% cut, so competitive pressure is a reasonable additional factor, though nobody outside OpenAI can confirm the split.
What is Fast mode in the OpenAI API? Fast mode replaced Priority Processing on July 30, 2026. For GPT-5.6 Sol it delivers up to 2.5x faster speeds than standard processing at twice the price, with no change in model intelligence. It is backward compatible: requests tagged with service_tier priority keep working and automatically use Fast mode, and you can pass either "priority" or "fast". The rename shipped inside the price-cut announcement, not its own release note.
How hard is it to switch between OpenAI and DeepSeek? Low, at the transport layer. DeepSeek serves an OpenAI-compatible endpoint and an Anthropic-format endpoint at api.deepseek.com/anthropic, and V4-Flash natively supports the Responses API and is adapted for Codex. For many applications, switching is a base URL and model name change. Output quality, latency and tool-calling reliability are separate questions a config change does not answer, and we did not benchmark them here.
Should I migrate to DeepSeek to save money right now? Only after running the numbers on your own workload shape, and only with the price warning built into the model. Input-heavy batch work may now be cheaper on OpenAI. Output-heavy and cache-heavy agent work is still meaningfully cheaper on DeepSeek, roughly 3 to 4x on realistic per-task math. Since a significant increase is coming with no date attached, treat current rates as temporary, not as a floor.
Sources
- OpenAI, "Advancing the price-performance frontier with GPT-5.6," July 30, 2026
- OpenAI API pricing rate tables
- Archived OpenAI pricing page, July 29, 2026, pre-cut rates
- DeepSeek Models and Pricing, price-increase notice at footnote 2, August 6, 2026
- DeepSeek Change Log, entry dated 2026-07-31
- Archived DeepSeek pricing, July 31, 2026 02:30 UTC, no pricing footnote
- Archived DeepSeek pricing, July 31, 2026 07:38 UTC, first with the surcharge footnote
- Archived DeepSeek pricing, August 5, 2026 09:56 UTC, last with the surcharge