DeepSeek bolted vision onto its cheapest model and charged nothing extra — the multimodal price floor just moved
TL;DR: On 21 August 2026 DeepSeek released deepseek-v4-flash-vision-exp, adding image understanding to its cheap workhorse V4-Flash — through the same API endpoint, with no price premium. Images bill by token count, capped at 384 tokens each, at V4-Flash’s existing rates; the new Files API is free. It is a sparse mixture-of-experts model, 13B active parameters of 284B total, with a 1M-token context. DeepSeek claims near-parity with Claude Opus 4.8 on multimodal agent benchmarks — ahead on Agents’ Last Exam (27.3 vs 25.7) and ZeroBench (35.0 vs 34.0), behind on ApexBench (36.5 vs 39.4) and Chartography (64.3 vs 65.0) — while trailing badly on text coding tasks like NL2Repo (57.7 vs 69.7). Every one of those numbers came from DeepSeek Harness Minimal Mode, which the company has not released. The model name ends in -Exp, and the release announces API access only, with no weights. So this is not a frontier substitution. What it is: the cost of evaluating a visual agent workload just fell by roughly an order of magnitude, and that is worth an afternoon of anyone’s time.
What shipped
DeepSeek’s release note is short and unusually specific about the commercial terms, which is the tell for what the release is actually for.
| Detail | |
|---|---|
| Model ID | deepseek-v4-flash-vision-exp |
| Architecture | Sparse MoE — 13B active of 284B total parameters |
| Context | 1,048,576 tokens; up to 384K output |
| Image input | base64, external URL, or Files API |
| Image cost | Token-billed, capped at 384 tokens/image, at V4-Flash rates |
| Files API | Free — upload once, reference by ID across requests |
| Endpoints | Chat Completions, Messages, Responses |
| Text ability | Retains V4-Flash’s agentic behaviour, reasoning, world knowledge |
The design decision worth pausing on is the 384-token cap per image. Most vision APIs scale image cost with resolution, so a high-detail screenshot can cost more than the prompt around it. A hard cap turns image cost into a constant you can multiply, which makes budgeting a visual pipeline arithmetic rather than estimation. At the off-peak cache-miss rate of $0.22 per million input tokens, an image tops out around $0.0001. A hundred thousand screenshots costs about ten dollars to look at.
The free Files API compounds it. If your agent examines the same image across several turns — a common pattern in UI automation and document review — you upload once and pay the token cost per reference rather than re-uploading.
The benchmark table, and its asterisk
DeepSeek published head-to-head numbers against Claude Opus 4.8 and against text-only V4-Flash.
| Benchmark | V4-Flash-Vision-Exp | Claude Opus 4.8 |
|---|---|---|
| Agents’ Last Exam (multimodal) | 27.3 | 25.7 |
| ZeroBench Pass@5 (multimodal) | 35.0 | 34.0 |
| Chartography (multimodal) | 64.3 | 65.0 |
| ApexBench (multimodal) | 36.5 | 39.4 |
| Terminal Bench 2.1 (text) | 83.9 | 85.0 |
| NL2Repo (text) | 57.7 | 69.7 |
| DSBench-Hard (text) | (≈8 pts behind) | — |
The shape is coherent and, to DeepSeek’s credit, not flattering everywhere. Multimodal results cluster within a couple of points either way. Text results split: Terminal Bench is close, but the twelve-point NL2Repo gap says this is not a coding model competing with a frontier flagship. That is exactly what you would expect from a vision layer grafted onto a cheap workhorse, and the fact the table shows it rather than hiding it is a small mark in its favour.
Now the asterisk, and it is a large one. DeepSeek states the evaluations were run using “DeepSeek Harness Minimal Mode” with specific parameters. That harness has not been released. None of these numbers are independently verifiable, and the comparison figures for Opus 4.8 were produced by DeepSeek running someone else’s model through its own scaffolding.
This is not a new complaint — it is the same reproducibility gap we flagged when DeepSeek’s pricing and flagship strategy shifted in mid-August, and the same one that applies to most vendor-run agentic evaluations, including Western ones. Agentic benchmarks are unusually harness-sensitive: the same model can swing many points on retry policy, tool definitions and prompt scaffolding alone. Treat the table as the vendor’s claim about the shape of the model, not as a measurement.
Why the pricing is the actual news
The benchmarks will be argued about. The price will not, because it is published and it is zero.
Set against DeepSeek’s own line-up, at published rates per million tokens:
| Model | Input (cache miss) | Output | Vision |
|---|---|---|---|
| V4-Flash-Vision-Exp | $0.22 – $0.44 | $0.66 – $1.32 | Yes |
| V4-Flash | $0.22 – $0.44 | $0.66 – $1.32 | No |
| V4-Pro | $0.66 – $1.32 | $1.98 – $3.96 | No |
(Lower figure off-peak, higher at peak. Peak windows: 01:00–04:00 and 06:00–10:00 UTC.)
The vision variant is identical to the text model on every line. DeepSeek is giving away the capability, which is a strategic choice rather than a technical inevitability — vision inference is not free to serve, and every other vendor charges something for it.
Two consequences follow.
The floor for experimenting with visual agents has effectively collapsed. The thing that has kept most teams out of screenshot-driven automation, chart extraction and document triage is not that the models could not do it — it is that per-image costs made a pilot expensive enough to need approval. A pilot that costs single-digit dollars does not need approval. It needs an afternoon.
It sharpens the question of what the premium buys. If a cheap model lands within a couple of points of a flagship on multimodal agent tasks — even on the vendor’s own harness — the burden shifts onto the expensive option to demonstrate the gap on your data. That is a healthy pressure and it is the same dynamic playing out at the top of the market, where OpenAI just cut its frontier model below Anthropic’s for a dated three-month window.
The caveats that should shape what you do
It is labelled experimental, and that is information. -Exp in a model identifier is a vendor saying the endpoint may be renamed, superseded or withdrawn without much notice. Fine for batch work you can rerun. Poor for a user-facing path with no fallback.
No weights. The release announces API availability; we found no accompanying weights drop. For a company whose reputation was built on downloadable models — and for buyers who choose Chinese models specifically to run them inside their own boundary — that is a real limitation, not a footnote. If self-hosting is the requirement, this release does not meet it. Compare with genuinely open-weight releases like Z.ai’s GLM-5.3, where the artefact is the product.
Surge pricing has a geography problem. DeepSeek’s peak windows land on European mornings, so “DeepSeek prices” for an EU team frequently means the higher number, not the headline one — a pattern we flagged when the peak-pricing policy first appeared. One thing improved this week: from 23 August 2026, off-peak rates apply across all of Saturday and Sunday (Beijing time), which makes weekend batch processing meaningfully cheaper and is worth scheduling around.
Data residency is a separate decision entirely. Sending images to a Chinese API endpoint carries governance implications independent of price or quality. Screenshots in particular tend to contain far more than the thing you meant to send — open tabs, customer names, internal URLs. Whatever your policy is, this is the release that will tempt someone to route production screenshots through it, so the policy should be explicit before the pilot rather than after.
What to do this week
Run a bounded evaluation, not a migration.
Pick one visual workload you already have — screenshot triage, chart-and-table extraction from PDFs, document classification, UI regression checks. Assemble 50–100 real examples with known-correct answers. Run them through V4-Flash-Vision-Exp and through whatever you use today. Compare accuracy on your own data.
That experiment now costs approximately nothing, which is the entire significance of this release. If the results clear your bar, deploy it where a wrong answer is cheap to detect and correct, and keep the incumbent where it is not.
The verdict
DeepSeek did not release a frontier model this week. It released a cheap model that can now see, and declined to charge for the seeing.
The benchmark claims deserve scepticism until someone reproduces them outside DeepSeek’s own harness, and the missing weights plus the experimental label mean this does not belong on a critical path yet. But the pricing decision is not a claim — it is published, verifiable, and it moves the floor. Visual agent work that was too expensive to prototype in July is now cheap enough to test on a whim.
For where each model earns its price, see our DeepSeek V4-Pro vs Claude Opus 4.8 breakdown, the best AI agent tools guide, and the DeepSeek profile. The comparison worth keeping honest is against Gemini 3.7 Flash, which is multimodal, supported, and priced at $0.75/$3.75 through year-end — cheap enough that “free vision” is a smaller advantage than it first sounds, and stable enough to build on.
Frequently asked questions
Does image input cost extra on DeepSeek V4-Flash-Vision-Exp?
No. DeepSeek's release note states images are billed by token count at the same rate as V4-Flash's existing pricing, capped at 384 tokens per image. There is no separate per-image fee and no vision surcharge. In practice that means an image costs at most 384 input tokens — roughly $0.0001 at the off-peak cache-miss rate of $0.22 per million. The associated Files API, which lets you upload an image once and reference it by ID across multiple requests, is free to use. The text side is unchanged at $0.22–$0.44 per million input (cache miss) and $0.66–$1.32 per million output, depending on whether you hit a peak window.
How does it compare to Claude Opus 4.8?
On DeepSeek's own published numbers it is close on multimodal benchmarks and clearly behind on text. Multimodal: ApexBench 36.5 vs 39.4, Agents' Last Exam 27.3 vs 25.7 (DeepSeek ahead), ZeroBench Pass@5 35.0 vs 34.0 (DeepSeek ahead), Chartography 64.3 vs 65.0. Text: Terminal Bench 2.1 83.9 vs 85.0, but NL2Repo 57.7 vs 69.7 — a twelve-point gap — and DSBench-Hard trailing by around eight points. The pattern is consistent with what the model is: a vision layer on a cheap workhorse, not a new flagship. The critical qualifier is that all of these were produced by DeepSeek using its own Harness Minimal Mode, which has not been released, so none are independently reproducible.
Can I self-host it? Are the weights available?
DeepSeek's release note announces API availability only — a model identifier, the endpoints it supports, and how to pass images. We found no weights drop published alongside it. That is a meaningful difference from the releases that built DeepSeek's reputation, where downloadable weights were the point, and it matters for anyone whose interest in Chinese models is specifically about running them inside their own boundary. If self-hosting is your requirement, this release does not serve it today. The experimental label suggests weights may follow if the variant graduates, but planning on an unannounced release is not planning.
What does the '-Exp' suffix actually mean for production use?
Treat it as a vendor telling you in the model name that the endpoint may change or disappear. Experimental identifiers historically get deprecated, renamed or superseded on short notice, and there is no stated support commitment attached to this one. That is survivable for batch and offline work, where a broken endpoint means a failed job you rerun against something else. It is a poor fit for a user-facing path with no fallback. If you do put it in production, put it behind a routing layer with a tested alternative, keep the model identifier in configuration rather than in code, and instrument for the case where the endpoint starts returning errors rather than assuming it will keep working.
Is DeepSeek still the cheapest option, given its August price rise?
For multimodal work specifically, this release makes it very cheap again — but the blanket 'DeepSeek is cheapest' assumption stopped being reliable this month and this does not restore it. V4-Pro rose sharply on 17 August and now runs $0.66–$1.32 input and $1.98–$3.96 output per million depending on the peak window, which puts its peak output above Gemini 3.7 Flash's promotional rate. What stayed at the floor is V4-Flash, and it is V4-Flash that just gained vision at no premium. Also note the surge-pricing geography: peak windows fall at 01:00–04:00 and 06:00–10:00 UTC, which overlaps European mornings, so an EU team paying 'DeepSeek prices' is often paying the peak ones. From 23 August 2026 the cheaper off-peak rate applies across all of Saturday and Sunday, which makes weekend batch work materially cheaper.
What should we actually do with this?
Run a scoped evaluation rather than a migration. Pick one visual workload you already have — screenshot triage, chart and table extraction, document classification, UI regression checking — assemble fifty to a hundred real examples with known-correct answers, and run them through V4-Flash-Vision-Exp alongside whatever you use now. The cost of that experiment is now low enough to be a rounding error, which is the whole significance of the release. Judge on your own accuracy numbers, not on DeepSeek's benchmark table, because the harness behind that table is not public. If it clears your bar, deploy it where a wrong answer is cheap to catch and correct, and keep the incumbent for anything where it is not.
Sources
- DeepSeek API Docs — DeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now Live (21 Aug 2026)
- DeepSeek API Docs — Models & Pricing
- OfficeChai — DeepSeek Releases V4-Flash-Vision-Exp, Matches Opus 4.8 On Some Multimodal Benchmarks
- Caixin Global — DeepSeek Enters the Multimodal AI Race with Experimental Vision Model (22 Aug 2026)
- Seeking Alpha — DeepSeek says V4-Flash-Vision-Exp comes close to Anthropic's Opus 4.8
- OpenRouter — DeepSeek V4 Flash Vision Exp model card
Related tool reviews
Questions or corrections? Email Pick Right. Want the full list? See all news.