Claude Mythos 5 vs. Grok 4.7: Drei unabhängige Zahlen, die nicht übereinstimmen

Artificial Analysis scored Grok 4.7 three times on the same afternoon and got three answers that point in different directions: 46 on its Intelligence Index — two points above Grok 4.6 — 56 on its Coding Agent Index, nine points clear, and 1,657 Elo on AA-Briefcase, 111 points clear. That spread is the most interesting thing about S​paceXAI's September 21, 2026 release, and it is also the reason this matchup against C​laude M​ythos 5 is worth writing down rather than waving through. Grok 4.7 is a generalist that has become a specialist: its gains are concentrated in long-horizon agentic coding and knowledge work, and are nearly invisible on the aggregate benchmark. C​laude M​ythos 5 has been a specialist from the day it shipped on June 9, 2026 — A​nthropic's restricted cyber and biology flagship, ASL-3 classified, refreshed as C​laude M​ythos 5.1 on September 1, priced at the M​ythos tier's $10.00 input / $50.00 output per million tokens, and available through a verification program rather than a rate card. One of these models you can buy. The other you can apply for. Everything else in this comparison is downstream of that sentence.

Was jedes Modell ist und wer es Ihnen erlaubt, es aufzurufen

Grok 4.7 is S​paceXAI's current flagship for coding, agentic tasks and knowledge work, succeeding Grok 4.6 at an identical price and an identical context window. The vendor writes its name as S​paceXAI on both the developer docs and the news page, while the API is still served from api.x.ai and the identifiers are still grok ones — worth knowing because searches for either name land on the same product. It ships live in Cursor and in the company's own G​rok Build terminal agent, and the vendor positions it as the default recommendation for everything except audio, image and video work. The claim the vendor makes is narrower and more specific than a general-purpose pitch: it is strong at long-running software engineering and at verifying its own work.

C​laude M​ythos 5 is the restricted SKU of A​nthropic's June 9 frontier generation, sharing underlying weights with the public Claude Fable 5 and differing from it only in safeguard level — the cyber and biology classifiers that stay switched on in the Fable model. A​nthropic classifies it ASL-3, the tier it reserves for capabilities that could meaningfully uplift cyberoffense, and rations access accordingly. Its reputation rests on red-team findings rather than a published benchmark table: a roughly 27-year-old undiscovered OpenBSD bug, an FFmpeg vulnerability that had survived about 16 years and five million automated scan hits, a browser exploit chaining four vulnerabilities to escape both renderer and OS sandboxes, and a Linux kernel privilege escalation.

• Context and output — 500,000 tokens with up to 450,000 output for Grok 4.7 vs 1M tokens with 128K maximum output for C​laude M​ythos 5.

• Price per million — $2.00 input / $6.00 output below 200,000 input tokens for Grok 4.7, rising to $4.00 / $12.00 above that line, vs the M​ythos tier's flat $10.00 / $50.00.

• Cache read — $0.50 per million for Grok 4.7, doubling to $1.00 past the 200,000-token boundary, vs $1.00 per million on the June M​ythos generation and $0.25 after the September refresh.

• Modalität — Text-, Bild- und Dateieingabe mit Textausgabe bei beiden.

• Reasoning control — configurable effort across four levels for Grok 4.7 vs adaptive thinking that cannot be disabled for C​laude M​ythos 5.

• Zugang — ein Self-Service-Schlüssel plus die eigenen Apps des Anbieters vs. einem Antrag über das Cyber Verification Program oder das Life Sciences Verification Program, mit einer 30-tägigen Datenaufbewahrungspflicht.

Those two price rows are where the usual reading goes wrong. On the sticker Grok 4.7 is a fifth of the M​ythos rate. On a long prompt it is just over a third, because the tier boundary lands at 200,000 tokens and this is a model sold on a 500,000-token window.

Die drei unabhängigen Zahlen, richtig gelesen

Start with the one that got the least attention, because it is the largest. On AA-Briefcase, the evaluator's board for long-horizon knowledge work, Grok 4.7 lands fourth overall, behind three C​laude entries, and ahead of Claude Opus 5 at xhigh effort — at roughly half Claude Opus 5's cost per task. Whatever else is true about this release, that is a price-performance result rather than a capability gap, and it is the first time a G​rok model has produced one on that board.

Die Zahl des Coding Agent Index gehört zu derselben Geschichte. Neun Punkte über Grok 4.6 auf der Coding-Agent-Achse gegenüber zwei Punkten im allgemeinen Index sind kein Rauschen; sie besagen, dass die Trainingsarbeit speziell in agentisches und Coding-Verhalten geflossen ist und dass das, was sich verbessert hat, nicht auf einen breiten zusammengesetzten Wert aus zehn Evaluierungen übertragbar ist. Der allgemeine Index ist die Zahl, die ein Leaderboard ausgibt. Die Coding-Agent-Zahl ist diejenige, die ein Team, das lange Tool-Schleifen laufen lässt, tatsächlich erlebt.

Then there is the disagreement between the vendor and the evaluator on the same benchmark, which is the part worth remembering every time you read a launch table. S​paceXAI's own Terminal-Bench 4.0 figure for Grok 4.7 is 38%. Artificial Analysis's independent run of the same benchmark put it at 33%. Five points, in the same direction, on one axis, for one model — and neither party is lying. Agentic benchmarks measure a model plus a scaffold, harness settings differ, and repeated runs move. That is what a five-point spread looks like when nobody is cheating. Hold it next to any single vendor number you see, including A​nthropic's.

Which brings this to the M​ythos side and the reason the comparison cannot be completed. A​nthropic publishes no benchmark table for C​laude M​ythos 5. The figures that circulate with the name do not all come from the vendor:

• ExploitBench, 78,0 % — Anthropics eigener Wert zum Juni-Launch, etwa doppelt so hoch wie die 40 %, die es für Claude Opus 4.8 angegeben hat. Vom Anbieter angegeben, nicht reproduziert.

• CyberGym Level 1, 83.8% — the number most often printed beside this model, and it does not originate with A​nthropic. It appears in competitor comparison material with its own framing interest.

• Terminal-Bench 4.0 — 42.0% for the June generation and 60.9% for C​laude M​ythos 5.1, per A​nthropic's September refresh material.

Notice what that third row does to the matchup. On the one benchmark both tracks have a figure for, C​laude M​ythos 5.1's 60.9% is far above Grok 4.7's 33% independently measured. Those are still not subtractable: one is a vendor number on the vendor's own harness, the other is a third party's, and they were run at different times on different configurations. But the gap is large enough that it is unlikely to be entirely an artifact, and it is a fair warning that a model with a general-purpose microphone is not competing on the axis the M​ythos tier exists for.

Weitschweifigkeit ist die Kostenzeile, die niemand ausdruckt

Grok 4.7's rate card is the cheapest thing about it and its verbosity is the most expensive. Artificial Analysis measured a 113% increase in reasoning-token use over Grok 4.6, with roughly 81,000 output tokens consumed per agentic task. Set that against a model that re-reads context rather than a model that re-derives it and the two cost structures stop looking alike at all: Grok 4.7 is cheap per token and expensive per task on long agentic work, while the M​ythos tier is expensive per token and its economics depend entirely on cache behaviour, which is why A​nthropic's September refresh cut the cache read from $1.00 to $0.25 and framed it as up to roughly 45% cheaper on highly agentic workloads.

Bei einem Preis von 6,00 US-Dollar pro Million entfallen auf eine Ausgabe von 81.000 Token für eine Aufgabe etwa 49 Cent. Bei dem 500.000-Token-Fenster, das dieses Modell anbietet, verdoppelt ein Prompt, der 200.000 Token überschreitet, den Eingabe-Tarif und damit auch den Cache-Tarif. Keine der beiden Zahlen ist ein Mangel — länger nachzudenken ist, wie Reasoning-Modelle besser werden —, doch eine Preisliste ist ein Preis pro Token, und Ihre Rechnung ist ein Preis pro Aufgabe, und der Multiplikator zwischen ihnen unterscheidet sich stärker, als es der Tarifunterschied tut.

Was der 6. Oktober verändert hat – und was nicht

On October 6, 2026, A​nthropic rebuilt the Cyber Verification Program into three application-based tiers, folded Project Glasswing into the top one, and turned C​laude M​ythos 5.1 from a partner-only allocation into an access grant any qualifying security organisation can apply for. Claude Opus 5.5 and C​laude S​onnet 5.5 joined every tier at the same time. Defense Access applications are answered within a few days per A​nthropic, Red Team Access takes a few weeks, and M​ythos access itself trails approval by roughly five business days. All of it vendor-stated and independently unconfirmed, and the bar remains a verification standard rather than a price.

What that means against Grok 4.7 specifically is narrow. For an authorised security team, this is the first time the M​ythos track has had a front door, and the comparison becomes real for the first time in this model's history. For everyone else — coding, knowledge work, long tool loops, the work Grok 4.7 is built for — the A​nthropic side of this matchup remains closed, and the fair reading is that it was never the right comparison to be making. The June 9 generation was suspended worldwide two days after launch by a US Commerce Department order and restored only for a limited set of vetted US organisations around the start of July; the model's whole existence has been shaped by that constraint, and no amount of benchmark reading changes who is allowed through.

Die Routing-Antwort, die heute die einzige verfügbare ist

Grok 4.7 is in the OrcaRouter catalogue at S​paceXAI's own rates — $2.00 input and $6.00 output per million, $0.50 cached, with the 200,000-token tier boundary and its $4.00 / $12.00 step-up carried through exactly as the vendor publishes them — passed at 0% markup, so a rate change is live here the day it posts. That precision matters more on this model than on most, because the cliff at 200,000 tokens is the single largest pricing discontinuity in this comparison and a router that flattens it into an average would mislead you about your own bill.

Das Zweite, was die Routing-Schicht leistet, ist, das Verbositätsproblem günstig testbar zu machen, statt es teuer anzunehmen. Eine durchschnittliche Ausgabe von 81.000 Tokens ist ein Durchschnitt, und Ihr Traffic kann durchaus deutlich über oder unter diesem Wert liegen. Grok 4.7 gegen Ihre eigene Arbeitslast laufen zu lassen, ist die einzige Messung, die beantwortet, ob es für Sie günstig oder teuer ist, und das ohne eine zweite Integration zu tun, bedeutet, dass die Antwort einen Nachmittag kostet. Automatisches Failover über Anbieter hinweg deckt die andere Hälfte ab – ein Reasoning-Modell, das unter Last länger nachdenkt, ist dasjenige, das am ehesten ein Timeout reißt, und eine Route, die die Anfrage an ein anderes Backend weitergibt, statt einen Fehler zurückzugeben, ist hier mehr wert als bei einem schnellen Modell.

If the capability you wanted from C​laude M​ythos 5 was the M​ythos class rather than the reduced safeguard level, the gated model cannot be routed to the public and no platform changes that. Its ungated twin can: Claude Fable 5.1 is the same underlying weights with safeguards left on, on the same catalogue at A​nthropic's list price of $10.00 / $50.00 with the September refresh's $0.25 cache read in place, and swapping between it and Grok 4.7 is a model-string change.

The verdict is narrower than the title suggests and it is worth saying plainly. For general agentic work, Grok 4.7 is the only one of these two you can call, and the independent evidence says its advantage over its predecessor is concentrated exactly where you would use it — which is a better sign than a broad two-point index gain would have been. C​laude M​ythos 5 is the stronger model on the one axis anybody has measured it against, and that axis is security work inside a vetted programme. If your work is not that, the interesting question is not M​ythos versus G​rok; it is whether Grok 4.7's per-task cost on your own traffic undercuts the model you are running now.

Nichts davon ist hypothetisch: Die dokumentierten KI-Ausfälle und Agent-Vorfälle, die hinter Releases wie diesem stehen, finden sich im OrcaRouter AI Incident Archive, wo jeder Eintrag seine eigenen Quellen, seinen Schweregrad, seinen Konfidenzgrad und sein Umstritten-Flag führt.