Two moves in twenty-four hours, and together they say more about where AI is heading than any model launch this month.

Thursday afternoon, OpenAI cut GPT-5.6 Luna's price by 80 percent and Terra's by 20 percent, three weeks after launch. Luna now runs $0.20 per million input tokens. Terra now costs less than Anthropic's mid-tier.

Early Friday morning, DeepSeek shipped the official V4-Flash API into public beta. Its small, cheap model now outscores its own larger flagship on every agent benchmark DeepSeek published, and it natively supports OpenAI's Responses API format with dedicated Codex adaptation.

Read that last part again. DeepSeek did not just cut prices to compete with OpenAI. It built its model to run inside OpenAI's own developer tooling.

Start with the result that has developers talking. DeepSeek-V4-Flash-0731 beats V4-Pro-Preview, the bigger and pricier model in the same family, on all nine agent and coding benchmarks DeepSeek listed.

Terminal Bench 2.1 went to 82.7, against 72.1 for V4-Pro-Preview and 61.8 for the Flash preview. DeepSWE jumped from 7.3 to 54.4. DSBench-FullStack went from 37.0 to 68.7. Cybergym landed at 76.7.

Here is the part that should stop you. The architecture did not change. V4-Flash-0731 is the same 284-billion-parameter mixture-of-experts model as the preview, identical size, identical structure. DeepSeek only redid the post-training. A sevenfold jump on DeepSWE came from better training after the fact, not from a bigger model.

For anyone who assumed capability gains require scale, that is a meaningful data point. The headroom in post-training is larger than most people were pricing in.

The benchmarks will get the attention. The integration is the actual move.

V4-Flash now natively supports OpenAI's Responses API format, and DeepSeek adapted it specifically for Codex. Per DeepSeek's setup guide, a single configuration exposes the model across Codex CLI, the ChatGPT desktop app, and the Codex extension for VS Code.

That makes V4-Flash a drop-in swap inside OpenAI's own ecosystem. Teams already building on OpenAI-format agent pipelines can point them at DeepSeek with minimal code changes, at reported input pricing around $0.14 per million tokens against Luna's freshly reduced $0.20.

It is a sharp piece of positioning. Rather than asking developers to leave OpenAI's tooling, DeepSeek made itself the cheaper engine inside it. Flash is currently the only DeepSeek model with documented Codex support, and V4-Pro compatibility is slated for early August.

DeepSeek did not fire the first shot. OpenAI cut Luna 80 percent and Terra 20 percent on Thursday, putting Luna at $0.20 and $1.20 per million tokens and Terra at $2 and $12, below Anthropic's Claude Sonnet 4.6 at $3 and $15. Sol, the flagship, was left unchanged.

OpenAI credited efficiency gains, and the source of those gains is genuinely odd: it tasked GPT-5.6 with optimizing its own runtime, and the model's rewrites of production GPU kernels cut serving costs 20 percent while improving token-generation efficiency by more than 15 percent. The model made itself cheaper to run. That arrived two days after 1,100 lab employees signed a letter asking Washington for tools to slow automated AI development, which is a coincidence worth noticing even if a kernel optimization is nowhere near what that letter is about.

OpenAI also replaced Priority Processing with a Fast mode for Sol, 2.5 times the speed at twice the price.

Every benchmark above came from DeepSeek, and the fine print matters.

DeepSeek ran the public Code Agent tasks using its own DeepSeek Harness in minimal mode, a framework it has not released yet, at max effort with top_p 0.95 and temperature 1.0. Two of the test sets, DSBench-FullStack and DSBench-Hard, are DeepSeek's own internal benchmarks that nobody else can run. Harness choice alone can move scores by double digits on identical tasks, which makes cross-vendor comparisons directional rather than definitive.

To DeepSeek's credit, it published those conditions openly rather than burying them, which is more than most vendors do. And the comparisons against its own preview build are apples to apples, so the internal improvement is real regardless of how it stacks against Opus 4.8 or GLM-5.2.

The honest position: the jump is genuine, the ranking is unverified, and independent numbers will land within days now that the API is public.

The stories moving fast beyond today's headlines:

Two moves in twenty-four hours reset the price war. On Thursday OpenAI cut GPT-5.6 Luna by 80 percent to $0.20 and $1.20 per million tokens and Terra by 20 percent to $2 and $12, putting its mid-tier below Anthropic's Claude Sonnet 4.6, while leaving flagship Sol unchanged and crediting efficiency gains produced by tasking GPT-5.6 with optimizing its own GPU kernels. Early Friday DeepSeek answered, putting the official V4-Flash API into public beta as build 0731, which keeps the identical 284-billion-parameter architecture as the preview and was only re-post-trained, yet beats the larger V4-Pro-Preview on all nine published agent benchmarks including 82.7 on Terminal Bench 2.1 against 72.1, and jumps from 7.3 to 54.4 on DeepSWE. The strategic move is protocol rather than price: V4-Flash natively supports OpenAI's Responses API and is adapted for Codex, so a single configuration exposes it across Codex CLI, the ChatGPT desktop app, and the VS Code extension, making it a drop-in swap inside OpenAI's own tooling at reported input pricing near $0.14 per million tokens. Every benchmark is DeepSeek's own, run on an unreleased in-house harness with two proprietary internal test sets, so the internal improvement is verifiable but the cross-vendor ranking is not. The takeaway is that switching costs, not benchmark scores, are becoming the real battleground, and the only number that settles it is cost per completed task on your own workload.

Stay building. 🤖