DeepSeek retired its old API names today, finishing the quiet handoff to V4

DeepSeek's old API aliases stopped working today. As of 15:59 UTC on July 24, the deepseek-chat and deepseek-reasoner endpoints that developers have called for more than a year are permanently retired, and every request against them now routes to DeepSeek-V4-Flash whether the caller updated their code or not. It's a small, mechanical deadline — the kind of thing that usually lives in a changelog — but it marks the real end of DeepSeek's slow migration off V3.2 and onto V4, a model line that shipped as a preview in April and has spent the months since quietly becoming the default.

V4 first appeared on April 24, 2026, as an open-weight preview in two sizes: V4-Pro, a 1.6-trillion-parameter mixture-of-experts model with 49 billion active parameters per token, and the lighter V4-Flash, at 284 billion total and 13 billion active. Both ship under the MIT license — no usage caps, no revenue thresholds, full commercial rights — and both default to a 1-million-token context window across DeepSeek's official services. At 1.6 trillion total parameters, V4-Pro is now the largest open-weight model publicly available, ahead of Moonshot AI's Kimi K2.6 and DeepSeek's own V3.2.

The architecture behind the efficiency claims

The headline spec is the context window, but the more interesting engineering is what makes a 1-million-token window affordable to actually serve. V4 alternates two attention mechanisms across its layers. Compressed Sparse Attention pools key-value entries down by a factor of four using a learned, softmax-gated compression, then uses a lightweight indexer to pick only the most relevant compressed blocks for each query. Heavily Compressed Attention goes further, compressing by a factor of 128 and attending densely across the whole compressed set rather than selecting a subset. DeepSeek reports that the combination lets V4-Pro handle a 1-million-token context using only 27% of the single-token inference compute and 10% of the memory that V3.2 needed for the same job.

On capability, DeepSeek's own benchmark writeup puts V4-Pro at 80.6 on SWE-bench Verified, the standard measure of whether a model can independently resolve real GitHub issues — within two-tenths of a point of Claude Opus 4.6's 80.8 in the same evaluation. On MCPAtlas Public, a tool-use benchmark, V4-Pro scored 73.6 against Opus 4.6's 73.8, a similarly narrow gap. Those are not wins, and DeepSeek doesn't frame them as such; they're a claim of parity on agentic coding work from a model whose weights anyone can download, fine-tune, and run on their own hardware.

A pricing structure built around Beijing business hours

When DeepSeek announced the mid-July timeline for V4's official launch on June 30, it paired the announcement with a new pricing structure that has no real precedent among the major model APIs: peak and off-peak rates, tied to the clock. API usage between 9 a.m. and noon and again from 2 p.m. to 6 p.m., Beijing time, is billed at twice the off-peak rate. It's a demand-shaping move that only makes sense if serving capacity, not model capability, is the binding constraint — the kind of pricing an electric utility runs, not a software company. For developers outside China, the practical upshot is that the same API call can cost twice as much depending on what time it lands relative to Beijing's workday.

What the cutover actually forces, and what it doesn't

It's worth being precise about what today's deadline does and doesn't confirm. DeepSeek has never published a single, formal 'V4 is now stable' announcement the way a lab like Anthropic or OpenAI might frame a general-availability date; the company's own timeline runs from an April preview to a vaguer 'mid-July' target for the official release, and outside observers still describe the exact graduation date as unconfirmed. What is concrete, because it sits in DeepSeek's own API documentation with a timestamp attached, is that the old names are gone as of today and every legacy call now resolves to V4-Flash. Functionally, that is the migration completing, whatever DeepSeek chooses to call the milestone. The move lands within days of Moonshot AI's Kimi K3, another massive open-weight release out of China this month — a reminder that the current wave of frontier-scale open weights is arriving from Chinese labs on a cadence Western labs, with their closed flagships, aren't matching.

None of this makes V4 a strict replacement for a closed frontier model, and DeepSeek's own numbers don't claim otherwise — near-parity on two agentic benchmarks is not the same as leading the field, and independent, third-party evaluation of V4-Pro's real-world reliability is still thinner than the evaluation history behind Opus 4.6 or GPT-5.6. What the cutover changes is the baseline anyone building on open weights has to compare against. A 1.6-trillion-parameter, MIT-licensed model with a working million-token context and coding scores within a rounding error of a closed flagship, priced at a fraction of it outside Beijing's peak hours, resets what 'good enough to self-host' means. That's the quieter story underneath a deadline that, on paper, is just two API aliases going dark.