DeepSeek V4 Goes Stable: Old API Retired, Peak Pricing Live

Key Takeaways

DeepSeek's V4 model family has completed its transition from preview to general availability. The company first announced DeepSeek-V4 as an open-sourced preview on April 24, 2026, according to the official DeepSeek API documentation. On June 30, 2026, DeepSeek said via a report from TechNode that the official, stable version of V4 would ship in mid-July 2026, replacing the preview designation carried since April.

The most consequential date for existing API users, however, is July 24, 2026, 15:59 UTC. That is when the legacy deepseek-chat and deepseek-reasoner API endpoint names were fully retired and became inaccessible, according to official DeepSeek API documentation. As of today, July 25, 2026, that deadline has just passed. DeepSeek-V4-Pro and DeepSeek-V4-Flash are now the only officially supported models on DeepSeek's API. Alongside the retirement, DeepSeek introduced a new peak-hour pricing mechanism, its first time-based surcharge, which took effect as part of this same transition.

Feature Overview

The V4 lineup consists of two models aimed at different workloads. DeepSeek-V4-Pro is a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active parameters per inference pass. DeepSeek-V4-Flash is a smaller, faster, and cheaper alternative, with 284 billion total parameters and 13 billion active parameters. Both are positioned as the default choices for any workload previously handled by the legacy chat or reasoner endpoints.

A one-million-token context window is now the default across DeepSeek's entire official service lineup, applying to both Pro and Flash. Both models also support dual Thinking and Non-Thinking modes, letting a developer choose between faster, lower-cost responses and slower, more deliberate reasoning within the same model family, rather than needing a separate endpoint name for each behavior.

On the integration side, both models remain accessible through an OpenAI ChatCompletions-compatible API and an Anthropic-compatible API format. For teams that already built against either of those API shapes, the practical change is limited to updating the model identifier, from the retired deepseek-chat or deepseek-reasoner names to deepseek-v4-pro or deepseek-v4-flash, while keeping the same base URL and request structure.

DeepSeek did not leave the migration to happen overnight. In the lead-up to the July 24 cutoff, calls made to the legacy deepseek-chat and deepseek-reasoner endpoint names were transparently routed behind the scenes to deepseek-v4-flash, using its non-thinking or thinking modes respectively. That routing bridge meant that applications referencing the old names kept working for months after V4 itself shipped, even though the underlying model serving those requests had already changed.

Usability Analysis

For developers who already had integrations pointed at the legacy endpoint names, the day-to-day experience of this transition has likely been uneventful until now. Because DeepSeek routed old endpoint calls to V4-Flash ahead of the hard cutoff, most applications did not need emergency changes when the preview and later the stable release shipped. The real usability test comes now: any code, script, or product still hardcoding deepseek-chat or deepseek-reasoner as the model string will start failing, since those names are no longer accessible after July 24, 2026, 15:59 UTC. Teams that have not yet updated their model identifiers to deepseek-v4-pro or deepseek-v4-flash need to do so immediately.

The new peak-hour pricing mechanism adds a fresh operational consideration that did not exist under the legacy pricing structure. During peak hours, defined as 9:00 to 12:00 and 14:00 to 18:00 Beijing time daily, API usage is charged at double the off-peak rate. Outside those windows, pricing stays at prior levels. This is the first time DeepSeek has implemented time-based surcharging on its API, a departure from the flat, low-cost pricing image the company built with earlier releases. For latency-tolerant or batch-style workloads, such as offline document processing or scheduled data pipelines, shifting execution outside Beijing business hours becomes a straightforward way to reduce costs. For applications that must respond to users in real time during those windows, the higher rate is simply an unavoidable cost of doing business during peak demand.

Pros and Cons

Pros

1. A single, unambiguous model lineup, V4-Pro and V4-Flash, replaces the older naming scheme and reduces confusion about which model actually serves a given request.
2. A one-million-token context window as the default for both tiers is a meaningful baseline for long-document and large-codebase use cases.
3. Compatibility with both OpenAI ChatCompletions and Anthropic-style API formats keeps migration friction low for teams already using either convention.
4. The pre-cutoff routing of legacy endpoint calls to V4-Flash gave developers a multi-month grace period rather than an abrupt breaking change.
5. Dual Thinking and Non-Thinking modes let developers trade off latency and depth of reasoning without switching models entirely.

Cons

1. Peak-hour pricing introduces a new variable into cost planning, since the same request can now cost twice as much depending on the time of day it runs.
2. Available reporting has not disclosed the exact off-peak or peak per-token dollar rates, which leaves budget forecasting less precise until DeepSeek publishes concrete figures.
3. Any integration that still references the retired deepseek-chat or deepseek-reasoner model names will now fail outright, since the transparent routing safety net ended with the July 24 cutoff.
4. Time-based surcharging is a new and unfamiliar mechanic for developers used to DeepSeek's historically flat pricing, adding a small layer of operational complexity.

Outlook

This transition marks a shift in how DeepSeek manages both its model lineup and its commercial terms. Consolidating around V4-Pro and V4-Flash as the sole official models simplifies the API surface going forward, and the one-million-token context default suggests DeepSeek intends long-context support to be a standard feature rather than a premium add-on. The introduction of peak-hour pricing is arguably the more structurally significant change. It suggests DeepSeek is managing infrastructure demand more actively than before, likely in response to usage patterns concentrated during Beijing business hours. Whether this time-based pricing model expands to future releases, or whether other providers follow with similar demand-based surcharges, is worth watching as the industry's pricing strategies continue to evolve.

Conclusion

DeepSeek V4's move to general availability is less about new capability and more about operational maturity. The model lineup itself, V4-Pro and V4-Flash, was already known since the April preview. What changed on July 24, 2026, is that the legacy interface finally closed, and a new, more complex pricing structure took its place. This release is most relevant to developers and businesses already running production workloads on DeepSeek's API. Anyone still referencing the old endpoint names needs to update immediately, and anyone running latency-tolerant workloads should evaluate whether shifting execution outside Beijing peak hours reduces costs under the new pricing rules.

Editor's Verdict

DeepSeek V4 Goes Stable: Old API Retired, Peak Pricing Live is a workable proposition that fills a clear gap, even if it doesn't fundamentally change the landscape.

The strongest case for paying attention is simplified, unambiguous model lineup with V4-Pro and V4-Flash replacing legacy naming, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, 1M-token context window as a default feature across both model tiers adds practical value rather than just headline appeal. The broader signal worth registering is straightforward: deepSeek's V4 family has moved from an April 2026 preview to a mid-2026 stable release, with the legacy endpoint names now fully retired as of July 24, 2026. On the other side of the ledger, new peak-hour 2x pricing multiplier adds cost unpredictability during Beijing business hours is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, exact off-peak and peak per-token pricing figures remain undisclosed in available reporting narrows the set of teams for whom this is an obvious yes.

For multi-model deployment teams, cost-conscious operators, and developers willing to evaluate beyond the major labs, the smart move is to track its trajectory and revisit once the rough edges are filed down. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.