DeepSeek Announces V4 Official Release for Mid-July 2026
The DeepSeek team announced on 30 June 2026 that the official, production-grade release of DeepSeek V4 is scheduled for mid-July 2026. DeepSeek V4 has been available in API preview since April 2026, but the mid-July date marks its graduation to a stable production release with additional feature optimisations and performance improvements over the preview version. The announcement simultaneously introduces a structural change to how DeepSeek prices API access: a time-based, peak and off-peak pricing model — the first time DeepSeek has applied surcharges tied to time of day rather than a single flat rate per token. Both developments are directly relevant to development teams that have integrated DeepSeek V4 during the preview period and to teams evaluating DeepSeek as a production inference provider.
The Two Models: V4-Pro and V4-Flash
DeepSeek V4 ships in two variants targeting different use cases. V4-Pro is the flagship: a Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion active parameters per forward pass, paired with a 1-million-token context window. The active parameter count means the per-token compute cost during inference is comparable to a 49 billion parameter dense model, despite the much larger total capacity — a deliberate design choice for delivering frontier-grade reasoning at lower inference cost than a comparably performing dense architecture.
V4-Flash is the speed-optimised variant: 284 billion total parameters, 13 billion active, with the same 1-million-token context window. V4-Flash targets high-throughput, latency-sensitive applications where the full V4-Pro capability is not required for the task. Both models support the 1M context length, making them suitable for long-document analysis, large-codebase reasoning, and extended multi-turn agentic pipelines without content chunking.
Peak-Time Pricing: The First Time DeepSeek Charges by the Clock
DeepSeek's new pricing structure introduces a doubling of API charges during defined peak hours each day. Peak periods run from 9:00 a.m. to 12:00 p.m. and from 2:00 p.m. to 6:00 p.m., with standard off-peak rates applied outside those windows. This is the first time DeepSeek has introduced time-based surcharging on its API and represents a meaningful change for production workloads that have been scheduled or budgeted against a flat per-token rate. Teams running inference pipelines that process documents, run code analysis, or perform batch operations at scale will need to evaluate whether their workload scheduling can shift to off-peak hours to maintain the low-cost profile that made DeepSeek V4 attractive during its preview period.
The peak-time pricing model is a demand management mechanism. As DeepSeek's API usage has scaled through the V4 preview period — driven by competitive pricing relative to OpenAI and Anthropic for comparable coding and reasoning performance — the time-based surcharge signals that DeepSeek's infrastructure is under peak-hour capacity pressure sufficient to warrant differential pricing for the first time in the company's history.
Legacy Model Retirement on 24 July 2026
The deepseek-chat and deepseek-reasoner model identifiers will be fully retired and inaccessible after 24 July 2026 at 15:59 UTC. API calls to these deprecated model names will return errors after that cutoff. Developers and teams currently calling these model names in their API integrations must migrate to the new V4 model identifiers before that deadline. DeepSeek has published migration guidance in its API documentation. The retirement deadline is hard: there is no grace period or soft deprecation window after 24 July. Any team with production workloads calling the old model names needs updated integrations tested and deployed before that date.
Migration Checklist for Teams
For teams currently using deepseek-chat: update the model parameter in API calls to the V4-Pro or V4-Flash identifier, validate response format consistency in downstream parsing logic, and run the updated integration against your existing evaluation benchmarks before promoting to production. For teams using deepseek-reasoner: the reasoning model has been replaced by V4-Pro's extended thinking capabilities, and the migration path requires checking whether the API response structure for reasoning traces has changed in the production V4 release.
What the V4 Release Means for Development Teams Using DeepSeek
Indian development teams have been among the highest-volume users of DeepSeek's API since the V3 launch in late 2025, drawn by its combination of frontier-class coding capability, 1M context length, and pricing substantially below what OpenAI and Anthropic charge for comparable performance. The official V4 release offers two immediate planning requirements.
First, the peak-time pricing change requires workload auditing. Batch operations, scheduled data processing pipelines, and any inference jobs with flexible timing should be evaluated for off-peak scheduling. Workloads with timing constraints — user-facing inference that cannot be deferred — should be budgeted against the doubled peak rate. For teams running large-scale document analysis or code generation pipelines, the difference between peak and off-peak scheduling could halve the monthly API cost of their DeepSeek usage.
Second, the model name retirement on 24 July 2026 is a hard deadline. API integrations referencing deepseek-chat or deepseek-reasoner must be updated and tested before that date. Given the 16-day window from the 8 July announcement date, teams should prioritise this change above other non-critical engineering work.
The Bottom Line
DeepSeek announced on 30 June 2026 that V4 officially releases in mid-July 2026, graduating from the April 2026 preview to a stable production release. Two model variants: V4-Pro (1.6T total parameters, 49B active, 1M context) and V4-Flash (284B total, 13B active, 1M context). The release introduces peak-time API pricing for the first time — charges double during 9:00 a.m. to 12:00 p.m. and 2:00 p.m. to 6:00 p.m. daily. Legacy model names deepseek-chat and deepseek-reasoner are fully retired on 24 July 2026 at 15:59 UTC. For Indian development teams relying on DeepSeek in production, the priority actions are: migrate model identifiers before 24 July, audit workload timing against peak pricing bands, and validate the V4 production release against existing evaluation baselines before promoting.
Frequently Asked Questions
When is DeepSeek V4 officially releasing and what are the two model variants?+
DeepSeek announced on 30 June 2026 that the official production release of V4 is scheduled for mid-July 2026, graduating from the April 2026 API preview. The release includes two variants: DeepSeek-V4-Pro, a Mixture-of-Experts model with 1.6 trillion total parameters and 49 billion active parameters and a 1-million-token context window; and DeepSeek-V4-Flash, a speed-optimised model with 284 billion total parameters and 13 billion active parameters, also with a 1-million-token context window. Both variants include additional feature optimisations and performance improvements over the V4 preview released in April 2026.
How does DeepSeek's new peak-time API pricing work?+
DeepSeek's official V4 release introduces the first time-based pricing model in the company's history. During peak hours — defined as 9:00 a.m. to 12:00 p.m. and 2:00 p.m. to 6:00 p.m. each day — API usage is charged at twice the standard off-peak rate. Standard rates apply outside those windows. Teams running scheduled batch operations or inference pipelines budgeted against DeepSeek's previous flat per-token rate need to audit workload timing against the new pricing bands. Workloads that can be shifted to off-peak hours should be, and those that cannot should be re-budgeted at the doubled peak rate.
When do deepseek-chat and deepseek-reasoner model names retire?+
The deepseek-chat and deepseek-reasoner model identifiers are fully retired and inaccessible after 24 July 2026 at 15:59 UTC. API calls to these deprecated model names will return errors after that hard deadline — there is no grace period. Developers and teams currently referencing these model names in their integrations must migrate to the new V4 model identifiers and complete testing before that date. DeepSeek has published migration documentation in its API docs. Teams have approximately 16 days from the 8 July announcement to complete the migration.
Why does DeepSeek V4 use a Mixture-of-Experts architecture and why does it matter for pricing?+
DeepSeek V4-Pro has 1.6 trillion total parameters but only 49 billion active parameters during any given inference pass. In a Mixture-of-Experts architecture, only a subset of parameters — the active parameters — is used for each token prediction, with a learned router selecting which expert weights to apply. This means the compute cost per token scales with the active parameter count (49B) rather than the total (1.6T), making it possible to deliver a model with large capacity and strong reasoning performance at an inference cost closer to a 49B dense model. This architecture is the primary reason DeepSeek V4 can maintain competitive pricing relative to dense frontier models from OpenAI and Anthropic with comparable benchmark performance on coding and reasoning tasks.
Written by
TechPillow Team
Sharing insights on technology, product development, and the Indian tech ecosystem.