Claude Sonnet 5.5 vs DeepSeek V4 Pro: 20 Index Points, and a Cache Read at $0.022

Put the two rate cards next to each other and the comparison looks closed before it starts. Claude Sonnet 5.5, generally available since September 28, 2026, costs $2.00 per million input tokens and $10.00 per million output tokens. DeepSeek V4 Pro, released April 24, 2026, costs $0.66 and $1.98. Output is the line that matters in agent work, and $10.00 against $1.98 is a gap of five to one — a dollar against twenty cents per hundred thousand tokens. Then look at the capability side, where Artificial Analysis puts Claude Sonnet 5.5 at 56 on Intelligence Index v4.3 and DeepSeek V4 Pro at 36, a twenty-point gap on the same harness. That is the whole tension: Deep​Seek's model sits at a fifth of the output price and roughly two-thirds of the measured capability, and the two facts do not resolve into a single recommendation without knowing what your workload actually does with a token.

What each model is, precisely

Naming is the first place this comparison goes wrong, so fix it here. DeepSeek V4 Pro as we list it is deepseek/deepseek-v4-pro, released April 24, 2026, a 1,048,576-token context model with a 384,000-token maximum output and text-only input. Artificial Analysis tracks a snapshot it labels DeepSeek V4 Pro 0813 — an August 13 build — and the figures quoted below come from that row. The Sonnet side is Anthropic's anthropic/claude-sonnet-5.5, text, image and file input, 1M context, 128K maximum output. If you are comparing a vendor page against an evaluator page and the numbers do not line up, check the build suffix before concluding that one of them is wrong.

DeepSeek V4 Pro supports text input only. Claude Sonnet 5.5 takes images and files as well. That is the first structural difference and it is not a marginal one: any pipeline that ingests screenshots, PDFs or diagrams cannot simply swap one for the other, because one of them cannot see.

The rate card, line by line

• Input — $0.66 per million for DeepSeek V4 Pro against $2.00 for Claude Sonnet 5.5; DeepSeek is 67% cheaper

• Output — $1.98 per million against $10.00; DeepSeek is 80% cheaper, the widest single gap in this comparison

• Cache read — $0.022 per million against $0.20; DeepSeek is 89% cheaper on the line that dominates long agent loops

• Cache write — $2.50 per million for the five-minute window on Claude Sonnet 5.5; DeepSeek V4 Pro's catalogue row carries no separate cache-write line

• Context — 1,048,576 tokens against 1,000,000; DeepSeek is marginally longer, a distinction with no practical consequence

• Maximum output — 384,000 tokens against 128,000; DeepSeek wins by a factor of three on the ceiling that constrains bulk generation

• Input modality — text only against text, image and file

• Reasoning — both use configurable reasoning effort rather than a fixed thinking budget, so depth is a request parameter on each

There is a scheduling detail on the DeepSeek side that a rate-card comparison will hide. Our catalogue row carries a timed-pricing window: between 01:00 and 04:00 UTC, and again between 06:00 and 10:00 UTC, the listed rates are multiplied by two. At those hours V4 Pro costs $1.32 input and $3.96 output — still cheaper than Claude Sonnet 5.5, by 34% and 60% respectively, but no longer by the margins the headline numbers suggest. Batch workloads that can be scheduled are worth scheduling, and workloads that cannot are worth pricing at the peak rate rather than the average. This is also the kind of thing that only shows up if you read the catalogue rather than the marketing page, which is the argument for putting every candidate model behind one endpoint instead of three vendor accounts.

Two capability numbers, one of them not independent

On the third-party side the ranking is unambiguous. Artificial Analysis scores Claude Sonnet 5.5 at 56 and DeepSeek V4 Pro at 36 on Intelligence Index v4.3, both in a maximum-effort reasoning configuration. The same evaluator puts Claude Opus 5 at 51 and Gemini 3.1 Pro Preview at 30, so V4 Pro's 36 places it below the previous Anthropic flagship and above Google's Pro tier — a respectable position for a model at a fifth of the price, and a long way from the top.

The cost-per-task reversal appears here too, and in this matchup it runs the other direction from the last one. DeepSeek V4 Pro costs $0.67 to evaluate on the index suite. Claude Sonnet 5.5 costs $7.60. That is not a typo and not a rounding artifact: the newer Anthropic model emits 410 million tokens across the suite against a median of 88 million, and the DeepSeek model's verbosity does not come close to closing a price gap of that size. On unconstrained open-ended tasks the cheaper model is cheaper by an order of magnitude in practice, not merely on the rate card.

The vendor numbers need handling with more care. DeepSeek publishes a τ²-Bench figure of 96.2 for V4 Pro, which is a strong number, but it is vendor-reported and τ²-Bench has since been removed from the Artificial Analysis index in its v4.3 revision — so it is a figure that cannot be independently placed on the same scale as anything Anthropic publishes. Anthropic's own launch table for Claude Sonnet 5.5 carries its own caveats, including a footnote that the Max effort setting scores lower than Xhigh on FrontierCode and a note that two of the benchmarks were run on a pre-release deployment with a structured-outputs bug. Put the two vendors' tables side by side and you have two marketing documents, not a ranking. The only figures in this article that belong on one axis are the two index scores and the two per-task costs, because one evaluator produced all four.

Why the output ceiling matters more than the price

The number in this comparison that gets the least attention is the one that changes architecture: 384,000 output tokens against 128,000.

Output ceiling is not a comfort feature. It decides whether a task is one call or a chain. Generating a large source file, emitting a complete document, producing a long structured report, translating a book chapter — anything where the artifact is bigger than 128,000 tokens — cannot be done in a single call to Claude Sonnet 5.5, and must be decomposed into a sequence with state carried between calls. Decomposition means more input tokens re-sent on every step, more cache reads, more orchestration code, and a new class of failure where the seams between chunks are where the errors live. A model at five times the price that removes an entire orchestration layer can be cheaper in engineering hours before it is cheaper in tokens, and in the other direction a task that fits in one call is a task where the ceiling never enters the arithmetic at all.

So the ceiling question is prior to the price question. If your longest artifact fits under 128,000 tokens, ignore this section and compare on cost per finished task. If it does not, price the pipeline you would have to build, not just the tokens.

Switching cost and the fallback problem

The migration in either direction is mostly about prompting rather than code, with one exception worth budgeting for.

Both models configure reasoning through an effort parameter, but they are different vendors' parameters, with different levels and different defaults. A prompt tuned to a high-effort Anthropic setting does not transfer to a DeepSeek effort setting as a like-for-like swap; it transfers as a re-measurement. Expect to spend a day per workload re-establishing quality before you trust a cost projection, on either side of the move.

The exception is vision. Claude Sonnet 5.5 reads images and files; DeepSeek V4 Pro reads text only. Any part of your pipeline that hands a model a screenshot, a scanned page or a rendered chart has no DeepSeek equivalent, so a full migration is only available to the text-shaped subset of your traffic. That is a dividing line to draw early, because it usually means the answer is not one model but a split.

Both are routable on OrcaRouter today — deepseek/deepseek-v4-pro and anthropic/claude-sonnet-5 are both live catalogue entries, priced at the provider's list with 0% markup, on one credential. That matters most in exactly this kind of comparison, where the deciding factor is not which model is better in the abstract but which model a given request should go to. A split that sends text-only bulk work to the cheap model and vision work to the expensive one is a routing rule, and it is far easier to express when both endpoints answer to the same key and the same failover policy. Claude Sonnet 5.5 itself is not yet a route on our platform, so today's version of that split pairs the Anthropic model with DeepSeek V4 Pro rather than replacing it.

The decision, stated as a rule

Send it to Claude Sonnet 5.5 when the task needs to see — images, PDFs, screenshots, rendered output — when the artifact is longer than a page of prose and you want a frontier model writing it, or when a twenty-point index gap is the difference between a draft a human has to rewrite and one they can ship.

Send it to DeepSeek V4 Pro when the workload is text in, text out, high volume, and tolerant of a lower ceiling on reasoning quality: classification, extraction, summarisation, bulk rewriting, code transformation with a clear specification, and anything where you would otherwise be paying ten dollars per million output tokens to move words around. The 89% cache-read discount makes long-context agent loops on this model structurally cheap in a way nothing else in the comparison is.

The honest verdict: this is not a capability race that DeepSeek is losing, it is a resource allocation decision that most teams get wrong in the direction of buying capability they do not use. If your traffic is text and your quality bar is "good enough to review" rather than "good enough to ship unread," V4 Pro at 20% of the output price and $0.022 cache reads is the correct default, and the twenty index points are a tax you are currently volunteering to pay.