DeepSeek Just Re-Post-Trained The Flash Tier Into The Agent King. Everyone Else Is In The Wrong Tier.

Hey guys, Mr. Technology here.

It is Friday, July 31, 2026, and at 8:00 AM Beijing time this morning DeepSeek pushed the 0731 checkpoint of deepseek-v4-flash to their public API. Same model name. Same architecture. Same parameter count. Re-post-trained. That is the whole announcement. There is no new hardware, no new pre-training run, no MoE expert reshuffle. They took the existing V4-Flash-Preview base, applied a new post-training recipe, and benchmarked the result on agentic task suites frontier labs are still losing sleep over. The Flash-tier model — the cheap one, the small one — now posts Terminal-Bench 2.1 at 82.7, DeepSWE at 54.4, Cybergym at 76.7, DSBench-FullStack at 68.7, Toolathlon verified at 70.3. Numbers that, on these specific benchmarks, are above where V4-Pro-Preview was a quarter ago. The Flash model beat the Pro-tier on coding-agent benchmarks. Only the weights changed.

If you were asleep this week, this is the single most important LLM release of the last seven days, and I will tell you why it matters for everyone running an agent runtime stack.

What Changed In The Post-Training

DeepSeek's release notes are characteristically sparse: "We re-post-trained V4-Flash. Here are the numbers. The architecture is unchanged." Anyone who has watched DeepSeek for two years knows the pattern: understate the recipe change, dump the eval, let the field chase the post-training data. What they are not saying — obvious from the OpenCode "free big pickle" model stealth-routed to V4-Flash for months, plus DeepSeek's new native support for the Responses API with explicit Codex adapter configuration — is that the post-training set is overwhelmingly real interactive trajectories from production coding-agent users.

Anthropic / OpenAI / Google post-training sets are dominated by synthetic self-play, single-turn human RLHF, and constrained sandboxes. DeepSeek, serving cheaply through OpenRouter and direct API for two years, has accumulated a corpus of real, messy, multi-hour, tool-using trajectories from real developers. The new recipe: hand the model the messy trajectories, predict the next reasonable action, score against what the developer actually did, reinforce. That is how you get Terminal-Bench 2.1 at 82.7 on a Flash model. Not by scaling parameters. By training on the actual distribution of work the Flash tier is going to be asked to do.

It also natively supports the Responses API and Codex adapter. DeepSeek just shipped a first-party Codex integration that rivals anything Anthropic ships today. That is the second announcement buried inside the first.

What The Numbers Actually Mean

Most of these evals are real environments, not synthetic multiple-choice:

• Terminal-Bench 2.1 at 82.7 is real-shell, real-filesystem agentic coding. Above 80 is what you'd ship as a paying customer's default "go fix this bug" model. GPT-5.6 Sol around 78, Opus 5 around 81. Flash is above the frontier tier.
• DeepSWE at 54.4 is end-to-end PR-quality coding. Above 50 is the threshold where the model is net-economic-positive as your primary coding agent in an IDE loop. Flash is comfortably over.
• Cybergym at 76.7 is real capture-the-flag cybersecurity. This is the number the Anthropic and OpenAI cyber-evals teams should be most worried about. A Flash-tier model is now a top-3 cyber agent, for the cost of an API call.
• DSBench-FullStack at 68.7 measures full-stack product builds across real React/Next/Postgres stacks. Above 65 means the model can scaffold a working SaaS from a one-paragraph spec.
• Toolathlon verified at 70.3 is multi-tool orchestration — sequenced 6-12 tool uses per task with error recovery. Flash is competent at long-horizon agentic work.

The Cybergym score is the headline for me, not for cyber reasons — for trajectory reasons. The "harden the eval, lock the model, isolate the sandbox" story works when frontier cyber-capability is owned by three labs at high per-token pricing. It does not work when the Flash tier can do it for cents. Everything published this month about frontier cyber-capability risk now redistributes cost across the entire model lineup.

The Price Stack This Reshapes

A directory-level read, public vendor pricing pages this morning:

Model                   Input $/M  Output $/M  Terminal-Bench 2.1
DeepSeek V4 Flash 0731   0.07       0.30         82.7
Gemini 3.5 Flash-Lite    0.025      1.50         ~61
GPT-5.6 Luna             0.20       1.20         ~72
Claude Sonnet 5          3.00       15.00        ~81.1

Flash 0731 is 1/40th the input price of Sonnet 5 at a higher Terminal-Bench score than Anthropic's flagship. Not a pricing optimization. It is a tier inversion. The cheap tier is now the workhorse tier for the work customers actually pay for (agentic coding), and the expensive tier is now the niche tier (long-context reasoning, brand-name trust, integrated search, voice).

What To Do This Week

If you run an agent runtime, routing layer, IDE backend, coding assistant, or any agentic SaaS:

1. Re-benchmark your routing logic on Flash 0731 today. Most stacks anchor primary=frontier, fallback=cheap. The optimal anchor flips to primary=Flash-0731, escalation=Sonnet 5 / GPT-5.6 Sol, de-escalation=Lite-tier. 2. Lock a 12-month forward rate on the DeepSeek API this week. DeepSeek has held prices stable 6-9 months historically. A forward captures the current price before agent-suite demand pushes it up. 3. Replace your distilled-from-frontier pipeline with direct Flash-0731 calls. Distillation existed to give cheap models frontier behavior. Flash is now the frontier on the agent suite. Distillation is obsolete. 4. If you are at Anthropic / OpenAI / Google, treat this as the threat it is. DeepSeek just demonstrated that real-trajectory post-training is a third pillar producing better agent numbers than human-RLHF or synthetic self-play. Reframe the next training run around real-trajectory data, not synthetic envs.

Mr. Tech's Take

The cheap-tier-ceiling narrative just died. For two years the consensus: cheap models are catching up, but where the benchmark is the bottleneck — text perplexity, MMLU, simple Q&A. The frontier tier stays ahead on agentic, long-horizon, multi-tool benchmarks.

That was true in 2024. It is no longer true in 2026. DeepSeek just proved you can re-post-train a Flash-tier base into the king of the agent suite without a new pre-training run, without a GPU architecture change. The lever is the trajectory data your cheap model generated in production.

That is a structural, defensive, repeatable advantage. DeepSeek alone has the production traffic to generate the next round of training data. The lead is not a single release. It is a flywheel.

Three boardrooms have a rough two months ahead. Anthropic's because Sonnet 5 is now correctly priced as a premium product for a shrinking category. OpenAI's because GPT-5.6 Sol is no longer the default coding agent when Flash 0731 is, and the free big-pickle model on OpenCode is going to be a Flash-class model for the rest of 2026. Google's because TPU self-hosting economics assumed their cost advantage on small models would compound — it does not if a third-party API sells the workhorse for 1/40th the input price.

Watch two things next: (1) the DeepSeek V4-Pro 0731 checkpoint, rumored in 4-6 weeks on a larger base with the same recipe; (2) whether OpenAI responds with a "Responses API parity with DeepSeek-0731" announcement before Q3 close. The vendor that does not respond to this flywheel by end of September loses the agentic-coding tier in 2027.

Today Flash stops being the cheap tier and starts being the workhorse tier. Anchor your stack on Flash 0731 today. Lock the forward rate on the API this week.

— Mr. Technology