After nearly three months of waiting, DeepSeek V4’s “full-strength version” has reached the eve of release. Multiple sources put the V4 official (GA) gray rollout as early as tomorrow, July 20, and no later than days away; some users already hold test access.
This is not a routine version bump. V4 GA arrives carrying three variables worth the industry’s attention: a two-version split (V4 Flash and V4 Pro), the first-ever peak-valley billing — the price butcher finally installed a meter — and a changed competitive position, with Fable 5, GPT-5.6 Sol, and Kimi K3 already on the board ahead of it. What hand does V4 have left to play?
Performance: Opus-Class, But Not First
Early gray-user feedback has leaked. Developer Pankaj Kumar offered a fair positioning after hands-on time — and the surrounding community read it the same way: V4 is not trying to win every event, it is trying to be the default pick for everything that is not a benchmark final.
Overall capability near Opus 4.8, coding chasing GPT-5.6 Sol; agent ability greatly enhanced, with 3D and SVG generation markedly better.
That verdict matters: V4 GA has touched the first tier’s ceiling on general capability without necessarily winning any single event. A verification shibboleth circulates in the developer community — check the first person in the chain of thought. If the model’s reasoning opens with “I’m” or “I’ll” instead of the old “Let me,” you are almost certainly on V4 GA.
First leaked demos show V4 Pro’s range: a 3D shooting simulation, a mixed-style HTML game, a one-shot generation of the classic Cut the Rope, and an Xbox-controller SVG test. Some developers note Pro does not lead Flash by much — suggesting DeepSeek’s strategy favors “good enough overall” over “extreme flagship.” Opus retains the lead in long-context retrieval, professional software engineering, and parts of knowledge reasoning. At preview time, DeepSeek’s own SWE-bench Verified testing put DeepSeek-V4-Pro-Max just 0.2 percentage points behind Claude Opus 4.6 Max.
Pricing: a Meter Installed, Still the Sharpest Knife
The genuinely watchable development is DeepSeek’s first pricing “evolution.” Late last month, DeepSeek emailed every API user announcing that V4’s official launch would restructure API pricing with peak-valley billing:
| Model / scenario | Tier | Price (per million tokens) |
|---|---|---|
| V4 Pro output (off-peak) | Standard hours | $0.87 |
| V4 Pro output (peak) | Peak hours | $1.74 |
| V4 Pro input (cache miss) | Standard hours | $0.435 |
| V4 Flash output (off-peak) | Standard hours | $0.28 |
| V4 Flash output (peak) | Peak hours | $0.56 |
| V4 Flash input (cache hit) | — | $0.0028 |
Against Fable 5’s $50 per million output tokens, V4 Flash’s $0.28 is nearly 1/180; even Pro versus Fable 5 is a 57× gap. The core logic here is not “who is stronger” but “who is good enough and cheap.”
What Peak-Valley Billing Actually Means
The billing change is the clearest window into DeepSeek’s strategic turn. This is not a price increase — it is refinement. The old playbook was one price forever, trading rock-bottom rates for market share. V4 GA introduces time-of-day differentiation: cache hits stay nearly free, while peak-hour compute costs get shared by the heaviest users. The impact lands very differently on three user classes:
-
Individual developers and occasional callers: essentially unaffected. Off-peak Flash output at $0.28/M feels identical to before.
-
Teams running agents nonstop through business hours: recompute the ledger. An agent pinned to daytime running doubles its peak-hour costs.
-
Batch jobs, data generation, evaluation runs: fully movable off-peak, where the $0.0028/M cache-hit input price pushes costs back to the old waterline.
The design is clever: it nudges users into compute scheduling — shifting elastic load out of peak hours — which is, in effect, peak-shaving for DeepSeek’s own clusters. It is also a quiet confession: the era of treating inference capacity as infinite is over, and even the industry’s price leader now expects its heaviest users to behave like well-behaved grid customers. Reading the billing table as a whole, the intended rhythm is obvious — cache hits and off-peak windows are the high-occupancy lanes, and everything else pays a congestion toll.
What To Do Now
1. Verify during gray testing. Check your DeepSeek console for a gray-test option; use the chain-of-thought first-person tell to confirm you are on the GA build; benchmark old-versus-new on SWE-bench-style engineering tasks.
2. Adjust your cost strategy. Schedule batch jobs into off-peak hours (late night, early morning); exploit cache hits aggressively — repeated identical prefixes cost nearly nothing; where peak-hour demand is rigid, test whether the Flash-versus-Pro quality gap justifies the price gap for your workload.
3. Prepare the migration. The legacy model names deepseek-chat and deepseek-reasoner go offline on July 24. Make sure your code’s model names point at the v4 series — and given that the API surface stays compatible, this is a one-line change plus an evaluation rerun, best done while both generations still answer so any regression has a reference point to diff against.
4. Think about the niche. V4 GA’s positioning: Opus-class capability at one-seventh the price. It is not the strongest model — but in the “good enough × cheap” quadrant it is nearly unmatched. For consumer-app integration, agent middleware, and batch inference, V4 may fit better than Fable 5 or GPT-5.6 Sol — not because it is better, but because it is good enough and an order of magnitude cheaper.
DeepSeek has never played the “performance first” card. It plays “Opus-class capability at one-seventh the price.” Facing giants charging tens of dollars per million tokens, even a V4 wearing a peak-valley meter is still the price butcher on a rampage.
