Artificial Intelligence

The Model Changed but the API Name Did Not

DeepSeek swapped the model behind deepseek-v4-flash on 31 July. Same architecture, same parameters, same endpoint. Its agent scores moved by multiples.

Artificial IntelligenceBy Servola Tech Desk2026-07-315 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

The Model Changed but the API Name Did Not

Key takeaways

DeepSeek replaced the model behind the deepseek-v4-flash endpoint on 31 July with a build carrying the same architecture and the same 284 billion parameters, and its software engineering agent score moved from 7.3 to 54.4. Anyone calling that endpoint changed models without changing a line of code.
There is no dated model identifier to call. DeepSeek's own documentation states the calling method is unchanged and the bare name reaches the latest version, so version pinning is not available to you and your own evaluation is the only version control you have.
The benchmarks that moved most are the ones measuring action rather than answers: Cybergym roughly doubled from 38.7 to 76.7 and DeepSWE rose more than sevenfold. If you authorised this model for shell or repository access, you authorised a weaker one.
Re-run your tool-access evaluation against the current build this week, and put a scheduled re-evaluation in the calendar rather than waiting for a release note. Set it before the next model in the family, V4-Pro, gets the same treatment.

One line in the changelog does all the work

A team that called deepseek-v4-flash on Thursday and called it again on Friday received answers from two different models. Nothing in their code changed, nothing in their configuration changed, and no error was raised. DeepSeek published the official release of its V4-Flash API on 31 July under the build name DeepSeek-V4-Flash-0731, and the documentation is admirably direct about what that means for callers: the model has been updated, the calling method remains unchanged, and the bare name reaches the latest version.

This was not concealed. The changelog is public, the benchmark table is published, and the build carries a date in its name. The problem is not disclosure. The problem is that the disclosure lives somewhere your runtime does not read. Every signal a production system uses to notice that its model has changed stayed constant across the swap: the model string in the request, the architecture, the parameter count at 284 billion total and 13 billion activated, the context window, and the price.

DeepSeek is explicit that this is a post-training change rather than a new model. Same architecture, same size, re-post-trained. That framing is accurate and it is also the reason the change is easy to underweight. A re-post-train sounds like a refinement. What it produced here was a different set of capabilities behind an unchanged identifier.

The numbers that moved are the ones about acting

Read the benchmark table by category rather than by row and a pattern appears. On Terminal Bench 2.1 the new build scores 82.7 against 61.8 for the preview. On NL2Repo it scores 54.2 against 39.4. On Toolathlon-Verified it reaches 70.3 from 49.7. On DeepSWE it goes from 7.3 to 54.4, an increase of more than sevenfold. On Cybergym it moves from 38.7 to 76.7, very close to double.

Each of those measures the model's ability to do something rather than to say something: run terminal commands, work across a repository, chain tools, find and exploit a weakness. The API also now supports the Responses format natively and has been adapted for Codex, which is a statement about where DeepSeek expects the model to be pointed. This is a build tuned to be given hands.

Why it matters: capability and blast radius are the same axis once a model has tool access. Most organisations decided what a model may touch by watching what it could do, and a great deal of that watching happened against a build that scored 7.3 on a software engineering agent benchmark. The permission granted on that evidence did not expire when the evidence did. If your policy allows this model shell access, repository write access, or anything network-facing, the reasoning behind that policy was calibrated on a model that is no longer answering.

The cheap tier now outscores the expensive one

There is a second consequence hiding in the same table. V4-Pro-Preview, the larger and more expensive model in the family, scores 72.1 on Terminal Bench 2.1. The new Flash build scores 82.7. The cheap tier now beats the premium tier of its own family on at least one agent benchmark, and it did so overnight, under a name that did not change.

Routing logic in most stacks is static and was written once: hard agentic work to the capable model, cheap or high-volume work to the fast one. That rule encoded a capability ordering that held when it was written. It no longer holds on every axis, which means some teams are now paying the premium tier to do work the cheap tier does better, and doing so without any signal that the ordering flipped.

For context on the field rather than the family, the new build sits at 82.7 on that benchmark against 81.0 for Z.AI's GLM-5.2 and 85.0 for Claude Opus 4.8. The useful takeaway is not the ranking, which will move again. It is that a gap of this size closed through post-training alone, on unchanged hardware requirements, in a single release.

What version control looks like when you cannot pin

Start by accepting the constraint. Providers that expose dated snapshot identifiers let you pin a build and migrate deliberately. DeepSeek's documented API surface offers the bare names, with the dated build recorded as a model version rather than something you can call. So the discipline has to sit on your side of the boundary, because there is no contractual or technical lever that holds the model still.

Practically, that means three things. Keep a small suite of your own tasks, twenty or thirty cases that represent what you actually ask the model to do, with recorded outputs and a stored date. Run it on a schedule rather than on rumour, because a schedule catches a silent swap and a rumour does not. And log the build name your provider reports alongside every production call you care about, so that when behaviour shifts you can tell a model change from a prompt change.

Then re-run the authorisation question, not just the quality question. Quality regressions are visible and annoying. A capability increase is invisible and pleasant, and it is the one that quietly widens what a compromised or confused agent can reach. The V4-Pro official release is coming, and DeepSeek has said so. The evaluation you schedule this week is the one that will be waiting when it lands.

Frequently asked questions

Can I pin my application to the previous DeepSeek build?

Not through the documented API surface. DeepSeek lists deepseek-v4-flash and deepseek-v4-pro as the callable model names and records DeepSeek-V4-Flash-0731 as the model version behind the first of those. The documentation states that the calling method is unchanged and the bare name reaches the latest version, so there is no dated identifier to request. The open-weights release of the same build can be self-hosted, which is the only route to a frozen version.

Does a post-training change really matter if the architecture is identical?

It matters more than the architecture in this case. The parameter count and structure are unchanged, so every capability difference in the published table came from training rather than from scale. DeepSWE moving from 7.3 to 54.4 and Cybergym from 38.7 to 76.7 are behavioural changes of a size that would normally accompany a new model, and they arrived under the same name.

We only use the model for text generation, not agents. Is this relevant?

Less urgently, but the pinning problem is the same. If your outputs feed a template, a downstream parser, or a compliance review, a build change can alter formatting, verbosity and refusal behaviour without notice. Keeping a small recorded task suite and running it on a schedule costs very little and is the only way to attribute a change in output to the model rather than to your own prompt.

The industry has spent two years teaching operators to pin dependencies and to treat an unversioned import as a defect. Model endpoints have quietly kept an exemption from that rule, and this release shows what the exemption buys: a sevenfold capability change delivered to production systems that were not asked and were not told at the point of use. You cannot make the provider hold the model still. You can make your own evaluation the thing that notices.

DeepSeekModel versioningAI agentsBenchmarksVendor riskTool access

The Agent Was Never the Hard Part

OpenAI's Presence deploys real-time voice and chat agents with policies, permission limits, escalation and simulations. The model was never the blocker - governance was, and now it is a product.

4 min read

Artificial Intelligence

Beijing Business Hours Now Price Your AI

DeepSeek's V4 API doubles in price during Beijing business hours and retires its old model names on 24 July. What that clock costs a European team.

4 min read

Law & Regulation

The EA Buyout Now Hinges on Brussels

The largest take-private in history has cleared EA shareholders and US antitrust. The last real gate is the EU - and not classic merger review, but the Foreign Subsidies Regulation, facing its first blockbuster test.

3 min read

← All articles