Claude Mythos 5 vs Gemini 4 Pro: The Name Is Wrong, the Access Gap Is Real

If you are comparing Claude Mythos 5 with Gemini 4 Pro, the first thing to know is that only one of those names belongs to a model you can buy. Google has not shipped anything called Gemini 4 Pro. The frontier Gemini 4 model in production is Gemini 4 Argon, announced on 30 September 2026 to a closed group of cyber defenders and opened worldwide on 13 October 2026 at an introductory $2 per million input tokens and $10 per million output. Claude Mythos 5, released on 9 June 2026, costs $10 and $50 for the same units and is not sold self-serve at all. So the honest comparison is not "which flagship is better." It is a gated specialist at five times the output price against an open-to-everyone generalist on a promotional rate that expires on 30 November 2026.

The naming confusion is worth clearing up rather than routing around, because it has already produced a shelf of comparison pages quoting specs for a model that does not appear in Google's API documentation. "Gemini 4 Pro" circulates as a community label for internal checkpoints and for the Argon release itself. Google's developer documentation lists gemini-4-argon. Every figure below for the Google side comes from that model.

What Google actually shipped

Gemini 4 Argon's headline change is not the input window, which Google has not formally published. It is the output ceiling: one million tokens in a single response, up from 64K in the previous generation. That is a different kind of capability from a large context window. A million-token output allows a model to write an entire migration, a full test suite, or a long document in one trajectory without stitching, and it changes what an agentic loop can return in a single step.

The launch benchmarks Google published, all vendor-reported and not independently reproduced:

• DeepSWE v1.1 — 77.9%, against 74.2% for Claude Opus 5.5 and 74.1% for GPT-6 Astra on Google's comparison.

• Vals Index — 68.9%, first place on Google's composite of finance, legal, tax and coding tasks.

• Zapier AutomationBench — 51.3%, first place.

• LVBench — 91.7% on long-video understanding.

• CWE-bench v1 — 68% on vulnerability remediation, tied for first.

• Harvey Legal Agent Benchmark — 19.6% all-pass.

The weak spots are equally specific and less quoted. Argon trails Claude Opus 5.5 by about nine points on Terminal-Bench 4.0, and trails GPT-6 Astra by roughly 10.5 points on FrontierSWE v2 and Terminal-Bench Science. On GDPval-AA it scores 56%, which one tracker places 14th. Artificial Analysis gives it an Intelligence Index of 53 — third, tied with GPT-6 Astra and behind Opus 5.5 at 58.

The price gap has a calendar attached

Argon's $2/$10 is an introductory rate. From 1 December 2026 it becomes $4/$20. The comparison most pages run today silently assumes the promotional number is permanent, which makes the gap against Claude Mythos 5 look like 5× on output when it is 2.5× after the change. Cached input is the other half of the story: Argon discounts cache reads by 95%, to $0.10 per million, while Mythos 5's cache reads are $1.00 per million. For an agent that re-reads a large repository on every turn, that ratio decides the bill more than the headline output price does.

Mythos 5 has one pricing advantage that rarely gets mentioned: batch requests are half price, and its 1M-token context is documented rather than inferred. Google has published the output limit but not the input window, and third-party trackers have filled the gap with figures Google has not confirmed. If your application depends on a specific input ceiling, that is a procurement question, not a benchmark question.

Two benchmark sheets that never meet

There is no shared harness between these two models in their public numbers. Google reports DeepSWE v1.1; Anthropic's Mythos 5 sheet leads with SWE-bench Pro at 80.3%, Terminal-Bench 2.1 at 88.0%, SWE-bench Verified at 95.5%, CyberGym at 83.8% and ExploitBench at 78.0%. Those are different tests, run by different labs, on different dates. Putting them side by side produces a table that looks like a comparison and is not one.

The only honest cross-model reference point is a neutral index. Artificial Analysis lists Gemini 4 Argon at 53. Claude Mythos 5 does not appear on that index at all — the closest entries are its sibling Claude Fable 5.1 and its successor Claude Mythos 5.1. For a model that has been available since June, the absence of an independent score is itself information: the access restrictions that keep Mythos 5 out of general circulation also keep it out of most evaluation pipelines.

Access is the deciding variable, not quality

Claude Mythos 5 cannot be bought with a credit card. On 6 October 2026 Anthropic restructured its Cyber Verification Program into three application-based tiers — Defence Access, Red Team Access, and a Specialised Access tier reserved for critical infrastructure — with the old Project Glasswing partners folded into the top tier. Applications are reviewed, US organisations only at launch, and all Mythos-class traffic carries a mandatory 30-day retention policy that rules it out for anyone with a zero-retention commitment. There is no sign-up page.

Gemini 4 Argon went the other way in stages. It launched to vetted defenders through the Fairwind Program, then opened to paid API customers on 13 October. If you are building something this quarter and you are not a security organisation with a review timeline, one of these models is a line item and the other is a project.

That asymmetry also reframes the benchmark argument. A model you cannot call has a quality score, not a quality advantage. The relevant question for most teams is which generally available model gives them the closest thing to Mythos-class behaviour. On the Anthropic side that is Claude Fable 5.1: the same underlying model with safeguards intact, $10/$50, and no verification programme. On the Google side it is Argon itself, which is broadly available but is not the same product as a classifier-free cyber model.

How to choose without a head-to-head you cannot run

If your work is authorised offensive security, red teaming, or malware analysis inside an organisation that can pass the review, Claude Mythos 5 is the specialist and the access process is the price of admission. If your work is long-output code generation, migration, or document synthesis, Gemini 4 Argon is cheaper today, cheaper again after the December repricing than Mythos 5 on output, and available without an application.

For everyone in between, the practical answer is not to pick either as a single dependency. Claude Fable 5.1 and Claude Fable 5 are both in the OrcaRouter catalogue at list price with no markup, which means Anthropic's rate card reaches you unchanged and a price change is live the same day. Gemini 4 Argon is not in our catalogue yet; the Gemini models we do route are the 3.x line, so a direct Google contract is still the route for Argon. Running the Anthropic side through one endpoint alongside the models you already use is the cheap way to test whether Fable 5.1 clears your bar before you commit to a verification programme for its unrestricted twin.

What would change this comparison

Two dates matter. On 1 December 2026 Argon's price doubles, which narrows the output gap to 2.5× and makes the cache-read difference more important than the list price. And Anthropic's three-tier programme is still new enough that approval timelines are vendor-stated rather than observed. If Defence Access turns out to take days rather than weeks, the access argument weakens and the specialist becomes a realistic option for a much wider set of security teams. Until then, the fair summary is that Google is selling a frontier generalist at a promotional price and Anthropic is selling a restricted specialist to a list of approved organisations — and those are different products before they are different scores.