DeepSeek V4’s formal release appears close, with circulating information suggesting the model could go live as early as tomorrow, or within the next few days at the latest. Some users have already been granted gray-release access to DeepSeek V4 (GA).
Two versions are now being referenced in testing: DeepSeek V4 Flash and DeepSeek V4 Pro.
Users are watching chain-of-thought wording for signs of GA access
One of the main questions around the so-called full-strength DeepSeek V4 is simple: how do users know whether they have been included in the rollout? Blogger AiBattle shared an unofficial way to check by looking at the model’s chain-of-thought, or CoT, phrasing.
If the model opens its reasoning in the first person with expressions such as “I'm” or “I'll,” rather than the older “Let me,” users have likely been switched to the V4 GA build.
Early developer feedback places V4 near Opus 4.8, with mixed views on competitiveness
Expectations around open-source AI releases remain high as Fable 5 and GPT-5.6 Sol continue to draw attention. Developer Pankaj Kumar, after testing V4, offered a measured assessment of the model’s early performance.
- He said the overall result is close to Opus 4.8 level, while coding ability is approaching GPT-5.6 Sol.
- Agent capability has improved sharply, and both 3D and SVG generation look materially better.
- On the same task, V4 needs more iteration rounds than Fable 5.
Kumar also said that, in the current product landscape, V4 is unlikely to beat the newly released Kimi K3, though he expects its price to be much lower. If that performance profile ends up paired with that pricing, the report argues, DeepSeek could produce another “DeepSeek moment.”
First demo wave has started circulating
The first DeepSeek V4 (GA) test demos have already begun to spread online, and outside reaction is split. Some viewers believe the model can already match Claude 5, while some developers say the Pro version does not lead the Flash version by a wide margin.
One example cited in the report is a 3D simulated shooting game generated by V4 Pro. Its core gameplay revolves around target practice with a siege crossbow vehicle, and the basic UI functions are already in place.
Another demo is an HTML game generated by the V4 release that blends elements of Minecraft and No Man’s Sky, with what the report described as relatively strong playability.
A version of the classic Cut the Rope was also said to have been produced in a single run by V4. Other demos mentioned include an Xbox controller SVG test and additional game-generation outputs.
DeepSeek is set to introduce peak and off-peak API pricing
Pricing remains the biggest variable in the V4 launch narrative. Leaks cited in the report point in the same direction: the model may not lead on raw performance, but it could come in at a much lower price.
According to the article, DeepSeek emailed all API users late last month and said the V4 formal release would arrive in mid-July. When GA goes live, the company will also adjust API pricing and introduce peak/off-peak billing for the first time.
For deepseek-v4-pro, the listed rate is $0.87 per million output tokens during regular hours and $1.74 during peak hours. Uncached input is priced at $0.435 during regular hours.
For deepseek-v4-flash, the pricing is lower: $0.28 per million output tokens in regular hours and $0.56 during peak hours, while cached input is priced at $0.0028.
The report frames this as a notable shift for a platform long associated with aggressive pricing. It says the change is unlikely to matter much for individual users, but teams that run agents continuously during working hours may need to recalculate costs. At the same time, cached-input pricing remains very low, which could keep batch jobs, benchmarking and data generation cheaper if they are moved out of peak periods.
Pricing still compares favorably with Fable 5, while older model names are set to retire
The article says V4 still stands out on price-performance when compared with Fable 5, which is listed at $50 per million output tokens.
It also cites DeepSeek’s own earlier preview-stage benchmark on SWE-bench Verified, where DeepSeek-V4-Pro-Max trailed Claude Opus 4.6 Max by only 0.2 percentage points while costing one-seventh as much.
That said, Opus still leads in long-context retrieval, professional software engineering and some knowledge-reasoning evaluations, according to the report.
Separately, the older model names deepseek-chat and deepseek-reasoner will be officially retired on July 24.
Based on the information currently circulating, DeepSeek V4 is unlikely to emerge as the top model across every category. Fable 5, GPT-5.6 Sol and Kimi K3 are already in the field. The central point of attention here is different: whether DeepSeek can again deliver Opus-level capability at a sharply lower price.