Verdict: Claude Opus 4.8 is still the stronger model on every benchmark they share — but DeepSeek V4 Flash 0731 gets close enough on agent and coding tasks that for most routine workloads, the price gap makes it the smarter default. At $0.28 per million output tokens versus Opus 4.8's $25 (an 89x difference), a team that routes 80% of traffic to Flash and reserves Opus for the hardest 20% typically cuts its inference bill by 70–85% with negligible quality loss on everyday tasks.
Last verified: 2026-08-04 · Best for hard reasoning: Claude Opus 4.8 · Best for cost-per-task: DeepSeek V4 Flash 0731 · Best open-weight self-host: DeepSeek V4 Flash 0731 (MIT license) · Pricing and benchmarks are volatile — re-check before committing.
What changed on July 31, 2026?
On July 31, 2026, DeepSeek promoted V4 Flash from preview to official public-beta release with a checkpoint called DeepSeek-V4-Flash-0731. The model string deepseek-v4-flash now points to this build automatically. What makes it striking is what did not change: the architecture is identical to the April 24 preview — 284 billion total parameters, 13 billion activated per token in a Mixture-of-Experts (MoE) configuration, 1M-token context window, and 384K maximum output. DeepSeek's changelog states plainly that the model "keeps the same model architecture and size as DeepSeek-V4-Flash-Preview, and was only re-post-trained" (DeepSeek API changelog, July 31, 2026).
What re-post-training means here is a reinforcement-learning pass focused on agentic capabilities — coding, tool use, cybersecurity, and multi-step task completion. DeepSeek's V4 technical materials describe the post-training paradigm as using GRPO (Group Relative Policy Optimization) to independently cultivate domain experts via supervised fine-tuning and reinforcement learning, then unifying them through online policy distillation (DeepSeek V4 model card, Hugging Face). The result: the same size model, with materially higher agent scores.
How big was the benchmark jump?
DeepSeek published a nine-benchmark comparison table with the 0731 release. The most dramatic gain is on DeepSWE, a contamination-free agentic coding benchmark from Datacurve that uses hand-written tasks across 91 repositories in five languages to prevent models from recalling solutions seen during pretraining (DeepSWE/Datacurve, May 26, 2026).
DeepSeek V4 Flash 0731 agentic benchmark scores:
| Benchmark | Flash 0731 | Flash Preview | Gain | Source |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 82.7 | 61.8 | +33% | DeepSeek changelog |
| DeepSWE | 54.4 | 7.3 | +645% | DeepSeek changelog |
| Cybergym | 76.7 | 38.7 | +98% | DeepSeek changelog |
| Toolathlon (verified) | 70.3 | — | — | DeepSeek changelog |
| NL2Repo | 54.2 | 39.4 | +38% | DeepSeek changelog |
| DSBench-FullStack | 68.7 | 37.0 | +86% | DeepSeek changelog |
| Agent Last Exam | 25.2 | — | — | DeepSeek changelog |
The DeepSWE jump from 7.3 to 54.4 is the headline. That is a 645% improvement on the same architecture — achieved entirely through re-training, not a larger model or longer context. DSBench-FullStack nearly doubled from 37.0 to 68.7.
Flash 0731 also outscored DeepSeek's own larger V4-Pro preview on every published agent benchmark, despite having roughly one-sixth the total parameters (284B vs 1.6T) and roughly one-quarter the active parameters (13B vs 49B).
Important caveat: All of these agent scores are vendor-reported. DeepSeek tested using its own "DeepSeek Harness" in minimal mode with max effort, which is not yet open-sourced. The changelog itself notes this. Independent evaluation from Artificial Analysis placed Flash 0731 at 50 on their Intelligence Index — a solid score but not frontier-leading (Artificial Analysis, July 31, 2026). Until third-party labs reproduce these numbers on their own harnesses, treat the biggest jumps as directional, not confirmed.
Does DeepSeek V4 Flash 0731 actually beat Claude Opus 4.8?
No — not on a single shared benchmark. Claude Opus 4.8 wins all of them, with an average margin of about 5–6 points.
Here is the head-to-head on the benchmarks both models report:
| Benchmark | DeepSeek V4 Flash 0731 | Claude Opus 4.8 | Gap | Source |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 82.7 | 85.0 | −2.3 | DeepSeek model card |
| DeepSWE | 54.4 | 58.0 | −3.6 | DeepSeek model card |
| Agent Last Exam | 25.2 | 25.7 | −0.5 | DeepSeek model card |
| DSBench-FullStack | 68.7 | 71.6 | −2.9 | DeepSeek model card |
| NL2Repo | 54.2 | 69.7 | −15.5 | DeepSeek model card |
| SWE-bench Pro | 73.7 (self-reported) | 69.2% | +4.5 | DeepSeek tech report / Anthropic |
Opus 4.8 also leads on broader capability benchmarks that Flash does not vendor-report: 88.6% on SWE-bench Verified, 83.4% on OSWorld-Verified (computer use), 49.8% on Humanity's Last Exam, and 1890 on GDPval-AA (knowledge work) (Anthropic, "Introducing Claude Opus 4.8," May 28, 2026).
The story is not that Flash beat Opus. It is that the gap narrowed to a few points on several agent benchmarks — while the price stayed 1/50th to 1/90th of Opus depending on the token mix.
How much cheaper is DeepSeek V4 Flash 0731?
The per-token price difference is the single most important number in this comparison.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Cache-hit input | Source |
|---|---|---|---|---|
| DeepSeek V4 Flash | $0.14 | $0.28 | $0.0028 | DeepSeek pricing page |
| DeepSeek V4 Pro | $0.435 | $0.87 | $0.003625 | DeepSeek pricing page |
| Claude Opus 4.8 | $5.00 | $25.00 | $0.50 | Anthropic pricing |
On output tokens alone, Opus 4.8 costs $25/M versus Flash's $0.28/M — an 89x difference. On a blended 3:1 input-to-output ratio (typical for agent workloads), the blended cost per million tokens works out to roughly $10 for Opus 4.8 versus $0.175 for Flash — approximately a 57x difference. A commonly cited "~1/54" ratio falls within this range.
There are two nuances that affect the real cost picture:
1. Verbosity matters. Artificial Analysis reported that Flash 0731 consumed about 210 million output tokens on their evaluation suite versus a median of roughly 62 million for comparable models — meaning Flash generates 3–4x more tokens per task than the field median. Cheap per-token rates do not automatically mean cheap per-task results. Retries and failed runs bill in full (Artificial Analysis evaluation, July 31, 2026). The honest unit of cost is cost per accepted result, not price per million tokens.
2. Flash has 5x the concurrency. DeepSeek's documentation lists V4 Flash at 2,500 concurrent requests versus V4 Pro's 500. This matters for agent fleets running parallel work, and it holds regardless of how the benchmark story resolves.
Which tasks should you route to each model?
The practical question is not "which model is better?" — it is "which model for which task?" Here is a routing framework based on the verified benchmark data and pricing.
Route to DeepSeek V4 Flash 0731 when:
- You run high-volume classification, sorting, or triage tasks (lead sorting, message routing, category tagging).
- You need a coding agent for well-specified, single-repository changes where the scaffold handles the exploration.
- Cost-per-task is the dominant metric, not peak quality on the hardest tasks.
- You want open weights (MIT license) to self-host, fine-tune, or run fully offline.
- You need 2,500 concurrent requests for batch agent pipelines.
- Your context fits within 1M tokens and you benefit from automatic cache hits ($0.0028/M input on cached segments — 98% cheaper than cache miss).
Route to Claude Opus 4.8 when:
- The task is a hard, multi-step agentic problem requiring exploration across unknown codebases or long reasoning chains.
- Quality on the hardest 5% of tasks is worth 89x the output cost (complex SWE-bench Pro-class work, OSWorld computer-use tasks, complex tool orchestrations).
- You need the strongest available instruction-following on ambiguous or highly-nuanced prompts.
- You are building workflows that leverage Claude Code's Dynamic Workflows feature (spawning hundreds of parallel subagents for convergent reasoning) — an Opus 4.8 capability Flash does not have.
- Verifiability and trust in a proprietary, safety-tuned model matters for a business-critical output.
- You benefit from Opus 4.8's MCP-Atlas score of 82.2% (best-in-class tool use) — structured multi-step API orchestration.
Split routing (the default recommendation for most teams):
Route 70–80% of inference traffic to Flash for routine workloads and reserve Opus 4.8 for the top 20% of tasks by difficulty. Use OpenRouter, a custom routing layer, or a prompt-based escalation rule (start on Flash, escalate to Opus if confidence is low or the task fails). This is the pattern that captures most of the quality at a fraction of the cost.
Is the open-weights story different from the API story?
Yes, and this matters if you self-host.
DeepSeek shipped the original V4 Flash preview as MIT-licensed open weights on Hugging Face on April 24, 2026. As of the July 31 release date, the 0731 checkpoint weights were not yet uploaded to Hugging Face — the existing repo still serves the April preview checkpoint. Multiple independent sources confirm this gap (Artificial Analysis, v4flash.com, digitalapplied.com, all reporting July 31, 2026). If you download and run the free Hugging Face version today, you are getting the older, weaker preview model, not the 0731 build.
This means there are currently three distinct access paths, and they are not the same:
- Hosted API (
model="deepseek-v4-flash") — serves the 0731 checkpoint. This is what DeepSeek's benchmark scores come from. - Hugging Face weights (
DeepSeek-V4-Flash) — still the April preview checkpoint. Weaker agent scores. - Website/chat app — DeepSeek's announcement states the app/web models remain unchanged for now; do not evaluate 0731 by chatting on the website.
Treat these as three separate claims until DeepSeek publishes the 0731 weights or clarifies the relationship.
What is V4 Pro and when does it arrive?
DeepSeek has not yet released the official version of V4 Pro — the larger 1.6T parameter / 49B active flagship model. The changelog says it will follow "soon" without a specific date. If the same re-post-training technique that lifted Flash from 7.3 to 54.4 on DeepSWE is applied to the Pro architecture, the jump could be even more significant — but there are no published numbers yet.
For a deeper look at the V4 Pro preview's architecture and the Qwen family for comparison, see Qwen 3.8 Max: What Alibaba's 2.4T Open-Weight Model Actually Delivers.
What this means for you
If you are a developer or small business running AI workloads in production, the practical takeaway is a routing decision, not a model loyalty decision. DeepSeek V4 Flash 0731 is now good enough and cheap enough that it should be your default for the bulk of routine agent and coding tasks. Claude Opus 4.8 remains the tool to reach for when a task is genuinely hard — complex multi-step reasoning, ambiguous exploration, or when the cost of a wrong answer exceeds the cost of a more expensive model.
The teams that win in 2026 are not the ones using the single best model. They are the ones routing traffic by task difficulty and paying 1/57th the price for 80% of their work.
For a complete setup walkthrough on running Flash 0731 as a coding agent, see How to Run DeepSeek V4 Flash 0731 as a Free Coding Agent in 2026. For a system-first approach to AI agent infrastructure (rather than chasing any single model), see How to Build a Multi-Model AI Coding Workstation in 2026 and System Over Model: How to Plug DeepSeek V4 Flash 0731 Into Any Agent Framework.
FAQ
Q: Is DeepSeek V4 Flash 0731 better than Claude Opus 4.8?
A: No. Claude Opus 4.8 scores higher on every benchmark both models report, by an average of about 5–6 points. Flash 0731's advantage is price: it costs roughly 89x less per output token and 57x less on a blended-token basis. Use Flash for volume, Opus for difficulty.
Q: How much does DeepSeek V4 Flash cost?
A: DeepSeek V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens (cache-miss rates). Cache-hit input costs $0.0028 per million — 98% cheaper than cache miss. Prices are from DeepSeek's official pricing page, verified July 2026.
Q: Is DeepSeek V4 Flash open source?
A: The April preview checkpoint is MIT-licensed and downloadable on Hugging Face. The 0731 build is available via the hosted API (model="deepseek-v4-flash") but the 0731 weights were not yet on Hugging Face as of July 31, 2026. If you self-host the current Hugging Face download, you get the weaker preview model, not the 0731 scores.
Q: What is DeepSWE?
A: DeepSWE is a contamination-free agentic coding benchmark from Datacurve, released May 26, 2026. It uses 113 hand-written tasks across 91 open-source repositories in five languages, designed so no model has seen the solution during pretraining. It has replaced SWE-bench Pro in the Artificial Analysis coding agent index.
Q: Can I use Flash 0731 for production workloads today?
A: Yes, via the hosted API. It is in public beta, not general availability. The calling convention is standard OpenAI-compatible at https://api.deepseek.com with model="deepseek-v4-flash". Flash supports 2,500 concurrent requests versus Pro's 500, making it well-suited for batch and parallel agent fleets.
Q: Will DeepSeek V4 Pro get the same upgrade?
A: DeepSeek has not announced a date for the official V4 Pro release. The changelog says it will follow "soon." If the same re-post-training technique is applied to the 1.6T-parameter Pro architecture, it could be a larger jump — but no numbers exist yet.
Discussion
0 comments