DeepSeek V4 Flash vs Gemini, GPT and Claude: Similar AI Scores, Vastly Different Token Prices

On Artificial Analysis’s cross-vendor index, DeepSeek V4 Flash scores 49.93 and Gemini 3.6 Flash scores 50.07. Google charges 10.7 times more for input and 26.8 times more for output.

Those are list prices per token. What a given task costs depends on how many tokens each model spends reaching an answer, which no price table shows — so read 27x as the gap on the output rate, not a promise about your bill.

That is the comparison worth making, and it is not the one you get by matching tier names. Matched on measured capability instead, DeepSeek’s peers are Gemini 3.6 Flash, GPT-5.6 Luna and Claude Sonnet 5 — not the models with “Flash” or “Haiku” in the name.

What does DeepSeek V4 Flash actually cost?

Three numbers, from DeepSeek’s own pricing page.

The model version is DeepSeek-V4-Flash-0731. It carries a 1M-token context window and a maximum output of 384K tokens, supports both thinking and non-thinking modes with thinking as the default, and has a concurrency limit of 2,500.

For reference, the tier above it — DeepSeek V4 Pro — costs $0.435 input, $0.87 output and $0.003625 on a cache hit, with a concurrency limit of 500.

Those Pro figures have a history worth knowing, because it shows how movable these numbers are. On 1 May 2026 the same page listed them as a 75% discount: $0.435 against a standard $1.74, and $0.87 against $3.48. Today the discount markers and the higher figures are gone and the reduced rates are simply the price. V4 Flash carried no such marker — $0.14 and $0.28 have been the plain price throughout.

Which models is it actually competing with?

Not the ones with matching names. That is the easy comparison and it is the wrong one.

Matching by tier label — DeepSeek’s “Flash” against everyone else’s “Flash”, “Luna” or “Haiku” — assumes vendors use those words to mean the same thing. They do not. To place these models on one scale we used the Artificial Analysis Intelligence Index , which runs its own evaluations across vendors on a single methodology, rather than each vendor’s self-reported benchmarks, which are not comparable with each other.

On that index, DeepSeek V4 Flash scores 49.93. Here is where it sits, with Gemini 3.5 Flash-Lite kept in the bottom row for contrast rather than as a peer:

Two corrections fall out of that, and both were in an earlier version of this article.

Gemini 3.5 Flash-Lite is not a peer. It sits 13.45 points below DeepSeek V4 Flash — a larger gap than separates V4 Flash from Claude Opus 5’s tier. Including it because its price was close was exactly the error of matching on the wrong axis.

Claude Haiku 4.5 is not a peer either. Artificial Analysis scores it at 29.58 with reasoning — 20.35 points below DeepSeek V4 Flash. Its non-reasoning figure of 23.71 is labelled on the same site as an estimate pending independent evaluation, so only the 29.58 is a measured score. It is a real model with a real score, and that score is nowhere near this comparison.

Anthropic is the awkward case, and worth setting out honestly. Two of its models bracket DeepSeek V4 Flash:

Sonnet 4.6 is closer on score. It is also more expensive than Sonnet 5 — $3 and $15 against $2 and $10 — while scoring six points lower, because Sonnet 5’s current rates are introductory and expire on 31 August. So the nearest Anthropic model by score is the worse buy on both axes, and the tables below use Sonnet 5, which is what anyone choosing today would actually pick.

At matched capability, what does each one cost?

As multiples of DeepSeek V4 Flash:

The Gemini row is the one to sit with. Gemini 3.6 Flash scores 0.14 index points above DeepSeek V4 Flash and costs 10.7 times as much on input and 26.8 times as much on output. Whatever that 0.14 is worth, it is not obviously worth 27x.

GPT-5.6 Luna is the genuinely competitive answer: 1.31 points higher for 1.4x input and 4.3x output. Claude Sonnet 5 buys 3.42 points at 14.3x and 35.7x — and after 1 September, at 21.4x and 53.6x.

The oddity inside DeepSeek’s own range

DeepSeek V4 Pro costs more than V4 Flash and scores lower on this index: 44.27 against 49.93, at $0.435 input and $0.87 output against $0.14 and $0.28.

That is a 5.66-point deficit for roughly three times the price. On this measure there is no reading in which Pro is the better buy, which is worth knowing before the name persuades anyone otherwise. It may well win on things the index does not capture — but on the index, it does not.

What the capability match does not settle

Three things, and they matter enough to state before anyone acts on the tables.

Reasoning effort is not held constant. DeepSeek V4 Flash, GPT-5.6 Luna and Claude Sonnet 5 are all measured at max; Gemini 3.6 Flash is measured at high, as its own model page states. Different settings, not a missing one. Scores taken at different effort settings are not strictly like for like — and effort drives output tokens, which is the expensive side of every price above. A model that reaches its score by thinking longer pays for it twice.

This is one index. Artificial Analysis is a single evaluator with a single methodology. It is the right kind of source for this question — the same harness across vendors, rather than four sets of self-reported numbers — but a different evaluator would order these models differently, and none of them measures your workload.

Index points are not tokens. A model that scores a point higher but needs half as many attempts to get a task right is cheaper in practice than the per-token table suggests. Cost per finished task is the number that decides a bill. Artificial Analysis does publish a version of it — a cost per Intelligence Index task, and output tokens per task, which fold reasoning tokens and cache behaviour into the total — but the figures sit in charts we could not extract reliably, so none are quoted here. That metric, not the rate card, is where this question is actually settled.

Two dated changes that are not in most comparison tables

Both are published by the vendors themselves, and both change the answer.

DeepSeek is introducing peak pricing at 2x. The footnote on its own pricing page reads that the service “will soon adopt a peak/off-peak pricing policy. During peak hours, prices will be 2x the regular prices, applicable to all billing items”, with peak hours given as 09:00–12:00 and 14:00–18:00 Beijing time daily. That is seven hours a day. The effective date is “subject to the official announcement” and had not been announced when this was written.

The consequence is worth spelling out. At 2x, V4 Flash costs $0.28 input and $0.56 output. Its output is still less than half GPT-5.6 Luna’s $1.20 — but its input becomes more expensive than Luna’s $0.20. The headline “cheaper on both sides” survives off-peak and does not survive peak.

Nor is DeepSeek the only one moving. The $0.20 and $1.20 used here are the figures published on OpenAI’s pricing page on 31 July 2026; that page carries no change log, so a comparison against an older capture of it may not reconcile, and the date this article was read is the only anchor we can offer.

Claude Sonnet 5 gets more expensive on 1 September 2026. Anthropic’s pricing page lists two rows for the same model: $2 input and $10 output per MTok through 31 August 2026, and $3 and $15 from 1 September. That is a 50% rise on both sides, published in advance and dated. That row is already in the comparison above, and it is what takes Anthropic’s output multiple from 35.7x to 53.6x.

One other thing on Anthropic’s page that surprises people: Claude Fable 5 is among its most expensive models at $10 input and $50 output, double Opus 5’s $5 and $25. It is not alone at that price — Claude Mythos 5, listed as limited availability, matches it — and the deprecated Opus 4.1 and retired Opus 4 are still listed higher at $15 and $75.

What these numbers do not tell you

Four things, in rough order of how much they move a real bill.

Cache mechanics are not comparable. DeepSeek prices a cache hit at $0.0028 and applies it automatically. Anthropic charges separately for writing to cache — on Sonnet 5, $2.50 per MTok for a 5-minute cache and $4 for an hour — before you get the $0.20 hit rate, all three rising by half on 1 September. Google charges $0.15 for cached context plus a storage fee of $1.00 per million tokens per hour. Three different billing models wearing the same word. The single “cheapest cached input” column above is the only piece of them that can be lined up.

Context tiers. OpenAI charges a different rate once a request crosses into its long-context band — GPT-5.6 Luna goes from $0.20/$1.20 to $0.40/$1.80. DeepSeek publishes one price for its full 1M-token window. For long-document work that difference compounds quietly.

The free tier. Google publishes a free tier for Gemini alongside the paid one, which no amount of per-token comparison captures.

List price is not the bill. Reasoning tokens you never see, cache-hit rates in the real world, retries, and per-request overheads all sit between the rate card and the invoice. We took that apart separately in what Opus 5, Fable 5 and GPT-5.6 Sol actually cost , and the reasoning there applies to every model here.

And one thing this article deliberately does not do: rank these models on quality. Price per token says nothing about how many tokens a model needs to get an answer right, which is the number that actually decides cost per finished task.

So is DeepSeek V4 Flash the cheapest?

On list prices, off-peak, yes — by a wide margin on output and a narrow one on input.

Off-peak and on output, nothing here is close to it. On input, GPT-5.6 Luna is within 43%, and once DeepSeek’s peak pricing takes effect, Luna is cheaper on input for seven hours a day. On cached input its peers charge 7.1x to 71.4x more today, and 7.1x to 107.1x once Sonnet 5’s introductory pricing ends on 1 September. That matters most for agent-style workloads that re-send the same context repeatedly.

The honest summary is narrower than the headline. DeepSeek V4 Flash is priced like a tier below the models it actually scores alongside, and its advantage is widest on output tokens and cache hits — which is where reasoning-heavy and context-resending workloads tend to spend, though not where every workload spends.

Sources

Each vendor’s own pricing page, all read on 31 July 2026.

Anthropic’s pricing documentation is reached through redirects from docs.anthropic.com and docs.claude.com; the final URL is the one cited. Vendor pricing pages change without notice — every figure here is as published on 31 July 2026.