Claude Mythos 5 vs Gemini 4 Argon: One Door Opened in Thirteen Days, the Other Hasn't Opened Since June

G​emini 4 Argon and C​laude M​ythos 5 were both announced as capability-above-the-flagship models with restricted first access, and the two have since diverged completely. G​oogle announced G​emini 4 Argon on September 30, 2026 through a Fairwind Program cohort of trusted cyber defenders, and on October 13 the model string gemini-4-argon moved to Stable on G​oogle's own paid tier — thirteen days from announcement to a self-serve key, with introductory pricing of $2 per million input tokens and $10 per million output standing. C​laude M​ythos 5 has been around far longer: A​nthropic released it on June 9, 2026 as the restricted sibling of Claude Fable 5, refreshed it as C​laude M​ythos 5.1 on September 1, and it still has no public endpoint, no self-serve tier and no published benchmark table — access runs through the Cyber Verification Program or the Life Sciences Verification Program and nothing else. That difference, not the benchmark spread, is what decides this matchup for almost every reader.

Two models with the same origin story and different endings

Both arrived the same way, which is what makes the divergence instructive. A​nthropic shipped C​laude M​ythos 5 as the restricted SKU of its June 9 generation, sharing underlying weights with the public Claude Fable 5 and differing only in safeguard level, and gave it to "a small group of vetted partners." G​oogle shipped G​emini 4 Argon first to the Fairwind Program — vetted government authorities, critical-infrastructure operators and software maintainers — and said plainly that broader access to paid API customers, G​oogle AI Ultra subscribers, developers and enterprises would follow. G​oogle then did the thing it said it would do.

A​nthropic did something different, and it is worth being fair about it. On October 6 it rebuilt the Cyber Verification Program into three application-based tiers, folded Project Glasswing into it, and turned C​laude M​ythos 5.1 from a partner allocation into an access grant any qualifying security organisation can apply for. That is a real widening of the door — the first one in this model's history. It is also still a door with a verification bar rather than a price, and it is aimed at security organisations specifically rather than at a general developer audience. G​oogle's path ends at a rate card; A​nthropic's ends at an application review.

What each one costs, and the caveat on the Argon number

• Input — G​emini 4 Argon at $2.00 per million tokens introductory vs C​laude M​ythos 5 at $10.00; a 5× gap.

• Output — $10.00 per million vs $50.00; the same 5×.

• Cached input — Argon at $0.10 per million, 95% off input; C​laude M​ythos 5.1 at $0.25 on the newer SKU and $1.00 on the June generation.

• Context and output — 1M tokens of context on both; Argon's output ceiling was raised to 1M from 64K, against 128K for the M​ythos tier.

• Modality — Argon takes text, images, video and files with text output; the M​ythos tier takes text, image and file input.

• Access — a paid API key since October 13 vs an application to A​nthropic's verification programs with a 30-day data-retention requirement.

The caveat is the word "introductory." G​oogle's $2 / $10 is a launch rate and G​oogle has said as much; a rate described that way can move, and the number to budget against is the one G​oogle posts the day you commit. The other caveat is arithmetic rather than policy: Artificial Analysis measures the cost to run its whole index suite at $2,407 for Argon — about $1.99 per index task — which is roughly a third of what the same suite costs on Claude Opus 5.5 — a reminder that list price and realised cost diverge once a model needs fewer tokens and fewer retries to finish a task. No equivalent figure exists for C​laude M​ythos 5, because no independent evaluator can call it.

The benchmark asymmetry, which is the honest headline

G​emini 4 Argon has been measured. C​laude M​ythos 5 has not, and not because anyone thinks it is weak.

Argon's independently published numbers: Artificial Analysis Intelligence Index 52.6 to 53 depending on effort configuration, at a measured $1.99 per index task and 48.9 output tokens per second with a 2.79-second first token. Vals Index first place at 68.90% (±0.97) on its GDP-weighted economic-impact board. Vendor-reported task scores, which G​oogle published with the September 30 announcement and which remain unreproduced: Terminal-Bench 4.0 at 57.1%, SciCode at 61.8%, Humanity's Last Exam at 57.1%, long-context reasoning at 79.7%, AutomationBench first at 51.3%, and an Omniscience figure of 42.4 — the largest single-task lead over a comparable open model in that set.

C​laude M​ythos 5's published record is a different animal. A​nthropic maintains no benchmark card for it, so every figure in circulation is second-hand. The CyberGym Level 1 score of 83.8% most often printed beside the name does not come from A​nthropic at all — it appears in a rival vendor's comparison, which has its own interest in how tight the race looks. The ExploitBench figure of 78.0% is A​nthropic's own, reported at the June launch against 40% for Claude Opus 4.8. Beyond that there are the operational claims: a 27-year-old undiscovered OpenBSD bug, a 16-year-old FFmpeg vulnerability, a browser exploit chaining four flaws to escape both renderer and OS sandboxes, and a Linux kernel privilege escalation from user to root.

Set those side by side and the boards are not comparable, which is the point. One model has an index score, a per-task cost and a leaderboard position; the other has a reputation and a set of case studies that nobody outside A​nthropic's partner list can reproduce. Argon also has a documented criticism the M​ythos tier does not: Bloomberg reported on announcement day that some G​oogle employees with direct access said the model does less well on real work than its benchmarks suggest — front-end design among the tasks named — and described it as affected by benchmark optimisation, which G​oogle disputes. That report is unverified and contested, and it is still a data point that exists for one side of this matchup and not the other.

The comparison neither vendor is making

G​oogle's launch material positions Argon against C​laude O​pus 5.5 and GPT-6 Astra on coding, and publishes no comparison with either Claude Fable 5.1 or C​laude M​ythos 5.1. A​nthropic's material positions M​ythos 5.1 against its own predecessor. Neither vendor has published a head-to-head, which means the matchup in this article's title is one both companies have avoided and that no independent lab can run, because one of the two models cannot be called by an evaluator.

What can be said is where the design briefs diverge. Argon's published comparison set includes Harvey's Legal Agent Benchmark, Vals Finance Agent v2 and an indirect-prompt-injection suite from Gray Swan — knowledge work, economic output, and injection resistance. The M​ythos tier's evidence is offensive and defensive security work. Those are different jobs, and a buyer choosing between them is usually choosing between a general frontier model that happens to be strong on documents and a specialist that is unusually strong on attack surface. The cost difference, 5× on list price, is the price of that specialisation, and it is only worth paying if the specialty is what you need.

What you can call today

Neither model is on OrcaRouter's catalogue, and that is not a temporary state a platform can fix — G​emini 4 Argon is reachable through G​oogle's own API at G​oogle's rate, and C​laude M​ythos 5 is reachable through a verification program or not at all. Listing one of them here as available would be a lie.

What is on the catalogue is the reachable member of each family. Claude Fable 5 is C​laude M​ythos 5's public twin — same underlying weights, full safeguard stack — at A​nthropic's list price of $10.00 per million input and $50.00 per million output with 0% markup, and Claude Fable 5.1 is the newer equivalent at the same rate with the cache read cut 75% to $0.25. C​laude O​pus 5.5 is there at $4.00 and $20.00, and A​nthropic's own launch material says it is comparable to C​laude M​ythos 5.1 in biology and cybersecurity — which, read plainly, is the vendor telling you the restricted model's edge has narrowed to the refusal behaviour that a verification program exists to unlock. Put those behind one endpoint with automatic failover and the "I cannot get the model I want" problem becomes a routing config rather than a research project: the work that needs the M​ythos-adjacent capability goes to Fable 5.1, the work that needs cheaper general reasoning goes elsewhere, and swapping either is a model-string change on the same key.

What to watch

Three signals, all specific. Whether G​oogle's introductory pricing survives contact with real traffic, which is the trigger for the cost argument to become actionable rather than theoretical. Whether A​nthropic publishes a M​ythos 5.1 benchmark table on the same set as O​pus 5.5 — the single most useful document the vendor could release, and the only way a buyer can value the 5× premium. And whether the Cyber Verification Program's application bar keeps falling, because the October 6 rebuild was the first time this model's access policy moved toward openness and the rate of that movement is the thing worth tracking.

Until those land, the fair summary is that Argon is the model with a price, an endpoint and someone else's ruler applied to it, and C​laude M​ythos 5 is the model with a reputation and an application form. Two of those three things are checkable, and the third one is not.

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.