Claude Mythos 5 vs DeepSeek V4 Pro: A 15× Price Gap, and Only One of Them Is Cheap at What It Is Good At

C​laude M​ythos 5 and DeepSeek V4 Pro are not competitors in the usual sense, and the numbers say so before any benchmark does. C​laude M​ythos 5 is A​nthropic's June 9, 2026 restricted cyber and biology flagship — refreshed as C​laude M​ythos 5.1 on September 1 — priced at $10 per million input tokens and $50 per million output on the M​ythos tier and available only through a verification program. DeepSeek V4 Pro is the open-weights-adjacent MoE flagship served from D​eepSeek's own API since April 2026 and refreshed in August, at $0.66 input and $1.98 output off-peak, roughly a fifteenth of the M​ythos rate with cache reads thirteen times cheaper, behind a key you can create in ninety seconds. One of these is a capability you apply for; the other is a line item you optimise. The comparison is worth making anyway, because the two models are bad at the same things and good at the same things, and the price gap is largest exactly where the gap in ability is smallest.

What each model actually is

C​laude M​ythos 5 is the restricted SKU of A​nthropic's frontier generation, sharing its underlying weights with the public Claude Fable 5 and distinguished only by safeguard level — the cyber and biology classifiers that stay switched on in the public model. A​nthropic describes it as the latest update to M​ythos Preview, with gains in cybersecurity, biology and healthcare benchmarks, released to "a small group of vetted partners." Its published benchmark history is thin and mostly comes from launch coverage rather than a maintained card, and the CyberGym Level 1 figure of 83.8% that circulates with the name does not originate from A​nthropic. Its known strength is offensive and defensive security work: the model's reputation rests on autonomous vulnerability discovery, and its access envelope is built around that — the Cyber Verification Program, and the C​laude Security scanner, which puts M​ythos 5 in front of real codebases for Enterprise customers without giving them free-form prompting.

DeepSeek V4 Pro is the opposite in almost every respect that matters commercially. It is a Mixture-of-Experts model with 1.6 trillion total parameters and roughly 49 billion active, a 1,048,576-token context window, a 384,000-token maximum output, text-only input, and a published rate card. It went out through D​eepSeek's API in April, was refreshed with the 0813 build on August 13, and is still served today at unchanged billing — a point worth stating explicitly because in September D​eepSeek announced the model's retirement, set a date, and then reversed itself eleven days later under developer pressure. The footnote on D​eepSeek's own pricing page now reads as a continuity commitment with no timeframe attached.

Rate cards, with the row most comparisons leave out

• Input — C​laude M​ythos 5 at $10.00 per million tokens vs DeepSeek V4 Pro at $0.66 off-peak, $1.32 inside the peak windows; a 7.6× to 15.2× gap.

• Output — $50.00 per million vs $1.98 off-peak, $3.96 peak; a 12.6× to 25× gap.

• Cache reads — $1.00 per million for the June M​ythos generation, $0.25 on C​laude M​ythos 5.1 and Claude Fable 5.1, vs $0.022 for DeepSeek V4 Pro off-peak; the line item that dominates long agent runs and where the gap is widest.

• Context and output ceiling — 1M tokens both; 128K maximum output on the M​ythos side against 384K for DeepSeek V4 Pro.

• Modality — text, image and file input on the M​ythos side vs text-only on DeepSeek V4 Pro.

• Access — a verification program and a 30-day data-retention requirement vs a self-serve API key.

The peak-window row is the one that bites. D​eepSeek's off-peak rate is half of peak, and the peak windows are 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays; everything else, including full weekends, counts as off-peak. A batch job that runs on a schedule rather than on demand can land in a peak window without anyone choosing it, and the effective input rate becomes $1.32. That is still a fraction of the M​ythos rate, so the conclusion does not flip — but a "25× cheaper" summary that ignores the window is describing a model you may not be calling. On the M​ythos side, note the direction of travel on caching: A​nthropic cut the cache read 75% between the two generations, which tells you the vendor knows that long agentic work re-reads the same context and that the cache line is where a heavy user's bill actually lives.

The benchmark comparison that cannot be made

Here is the honest problem with this matchup, and it is not a small one. A​nthropic does not publish a benchmark table for C​laude M​ythos 5, and D​eepSeek does not publish results against M​ythos 5. The two models have been measured on different things by different parties, and the figures that circulate for C​laude M​ythos 5 mostly come from third-party comparisons rather than from A​nthropic.

What exists is asymmetric in a useful way. D​eepSeek published a full card for the 0813 build, all of it vendor-reported and none of it independently reproduced: Terminal-Bench 2.1 at 87.9, DeepSWE at 62.7, CyberGym at 83.3, Humanity's Last Exam at 42.7 without tools and 60.0 with them, Toolathlon-Verified at 74.1, AutomationBench at 31.8. Independent evaluators then cooled the picture considerably — Artificial Analysis places the 0813 build at 36 on its Intelligence Index with a measured cost of $0.67 per index task, and its non-reasoning variant at 20. That is a real capability profile for $0.66 per million input tokens, and a long way below the frontier on general reasoning.

The place the two models actually overlap is security, and there the comparison is instructive in the wrong direction. D​eepSeek reports CyberGym at 83.3 for V4 Pro. The number most often printed next to C​laude M​ythos 5 on the same benchmark is 83.8 — and it does not come from A​nthropic, it comes from a competitor's comparison that has its own reason to frame the race tightly. Two vendor-adjacent figures a half-point apart on the same benchmark is not evidence that an open MoE flagship matches A​nthropic's restricted cyber model. It is evidence that neither number has been independently verified, and that the benchmark is being used by both sides as marketing infrastructure.

Where each one is the wrong tool

Put the two models against a task list and the split is clean enough to route on.

C​laude M​ythos 5 is built for security work and is closed to almost everyone who wants it. If your task is vulnerability discovery, exploit development, red-teaming against your own infrastructure, or security research that needs free-form prompting of a frontier model, the M​ythos tier is the one designed for that, and the price is not the obstacle — eligibility is. The same model's public twin, Claude Fable 5, carries the full safeguard stack, which means penetration testing, exploit generation and binary vulnerability scanning stay blocked; you are buying the capability tier, not the cyber permissions.

DeepSeek V4 Pro is built for volume. Text in, text out, a huge output ceiling, and a price that makes aggressive retry loops and long extraction pipelines economically mundane. Its profile is exactly what you want for classification at scale, structured extraction from documents that are already text, summarisation, first-pass triage and the formatting-heavy middle of an agent pipeline. What it is not is a model you send an image to: it has no vision, so a scanned PDF or a screenshot is not a cheaper input path, it is an excluded one. And its general reasoning index of 36 means the tasks that need a genuine multi-step plan are the tasks where the price advantage stops compensating.

The cost structure that makes routing the answer rather than a compromise

Because the strengths do not overlap heavily, the useful move is not choosing. Send the vision-bearing work and the hard reasoning to the A​nthropic side and the high-volume text work to D​eepSeek, and the saving on a mixed workload is usually large — but only if the two models live behind one key, because otherwise the second integration costs more engineering time than the difference in token spend.

That is the case for a routing layer rather than a procurement decision. OrcaRouter carries anthropic/claude-opus-5.5 at $4.00 in and $20.00 out and deepseek/deepseek-v4-pro at $0.66 and $1.98 on the same O​penAI-compatible endpoint, in a catalogue of 200-plus models, with provider list price passed through at 0% markup — which matters on the D​eepSeek side specifically, because D​eepSeek has repriced V4 Pro more than once and moved its off-peak windows, and a pass-through route follows the vendor's card instead of freezing a number into your contract. The routing DSL expresses the split in one place — cheap model for the easy branch, escalate when the answer's confidence or the task's shape says so — and automatic failover is worth more here than on a stable model: V4 Pro spent a week in September under an announced shutdown that was then withdrawn, and having a route that falls back to a second provider without a deploy is a cheap hedge against a vendor changing its mind.

The other honest use of the same catalogue is as a measuring device. The expensive-model-first habit is common and expensive; the only reliable fix is counting tokens on real traffic with both models, which is a config change when both are on one key and a project when they are not.

What would change this comparison

Two things. If D​eepSeek ships a successor that closes the long-horizon agent gap at anything like the current price, the split collapses into a price question and the answer is trivial — and D​eepSeek has made that move repeatedly, including a September release it says comprehensively surpasses V4 Pro on performance, cost and speed. In the other direction, if A​nthropic publishes a M​ythos 5 or 5.1 benchmark table on a standard set, the security half of this comparison becomes checkable rather than asserted, and buyers would finally be able to tell whether the $10 / $50 rate card is buying something D​eepSeek's $0.66 does not provide.

Neither has happened. Until one does, the practical position is that DeepSeek V4 Pro is the better value per token on the market for text-shaped work and is priced as if frontier reasoning were a solved commodity, and C​laude M​ythos 5 is a capability most teams cannot obtain and a price most teams would not pay if they could. Build on the one you can call; keep the model name in config so the other becomes a one-line change if it ever opens.

None of that is hypothetical: the documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.