How Opus 4.8 compares to Claude Mythos and GPT 5.5

Anthropic launched Claude Opus 4.8 this week, and Artificial Analysis reports that it narrowly tops the independent Artificial Analysis Intelligence Index at 61.4, compared with 60.2 for GPT-5.5 (xhigh). In Anthropic’s announcement for the model, Opus 4.8 looks mostly like an incremental upgrade over 4.7, but the company noted that its heavily-promoted Mythos model would be generally available in the “coming weeks.” While Opus 4.8 ships at the same base price as Opus 4.7, $5 input and $25 output per million tokens, Mythos Preview initially launched to Project Glasswing participants at $25 and $125 per million tokens, five times the Opus rate, with Anthropic committing $100 million in usage credits to cover the preview.

So how much more advanced is the Mythos tier, and where does the advantage actually live? Anthropic’s system cards point to a general-intelligence gap over Opus 4.8 that is modest and uneven, while the cyber gap is vast.

The Mythos gap is uneven

On Anthropic’s internal AECI index, Opus 4.8 lands at 155.5, between Opus 4.7 (154.1) and Mythos Preview (158.3), per the Opus 4.8 system card. Anthropic computed that on a small subset (n=11), so it reads as a few index points rather than a generational leap. Meanwhile, Mythos leads on the hardest software work, posting 77.8 to Opus 4.8’s 69.2 on SWE-bench Pro, a test of whether a model can resolve real GitHub issues in large codebases, and 59.0 to 38.4 on SWE-bench Multimodal, the same task with visual elements like screenshots and diagrams. The gap narrows on pure reasoning. GPQA Diamond, a set of graduate-level science questions written to be Google-search-proof, is a frontier tie at roughly 94, and the two models sit within a point of each other on USAMO, the US Mathematical Olympiad’s proof-based problems.

General capability

Benchmark	Mythos Preview	Opus 4.8	GPT-5.5
SWE-bench Pro	77.8	69.2	58.6
HLE, no tools	56.8	49.8	41.4
HLE, with tools	64.7	57.9	52.2
GPQA Diamond	94.5	93.6	93.6
BrowseComp	86.9	84.3 single-agent, 88.5 multi-agent	84.4
CyberGym	83.1	78.8	81.8
OSWorld / OSWorld-Verified	79.6 OSWorld	83.4 OSWorld-Verified	78.7 OSWorld-Verified
Terminal-Bench	82.0 on 2.0	74.6 on 2.1	78.2 on 2.1, or 82.7 on 2.0

Sources: Anthropic’s Claude Opus 4.8 system card for Opus 4.8 figures and Anthropic-reported GPT-5.5 comparisons; Anthropic’s Claude Mythos Preview system card for Mythos figures; OpenAI’s GPT-5.5 launch materials for GPT-5.5 rows omitted from the Opus 4.8 card. Benchmark versions and harnesses are not always identical. Mythos was benchmarked before GPT-5.5’s release, so Mythos figures were not originally reported head-to-head against GPT-5.5. Terminal-Bench and OSWorld rows should be read as directional because versions and variants differ.

Cyber is Where the leap lives

Anthropic’s own Opus 4.8 card states the model is much weaker than Mythos on cyber, and the internal evaluations are stark. With safeguards off, Mythos produced a full working exploit on 70.8% of Firefox targets against Opus 4.8’s 8.8%, and failed to score on only 23.3% of OSS-Fuzz targets versus 61.5% for Opus 4.8. 

Why not include GPT-5.5-Cyber? GPT-5.5 complicates Anthropic’s framing. Berkeley RDI’s ExploitGym, the best public cross-model test so far, showed a significant lead for Mythos. ExploitGym paired Mythos with Claude Code against GPT-5.5 with Codex CLI across 898 real-world vulnerability instances. Mythos produced 157 intended working exploits to GPT-5.5’s 120, yet GPT-5.5 led on kernel exploitation, 22 to 12. The UK AI Security Institute found GPT-5.5 edging Mythos on expert cyber tasks, 71.4% to 68.6%, within overlapping uncertainty. 

Cyber comparison, and how GPT-5.5 stacks up

Cyber evidence	Mythos Preview	Opus 4.8	GPT-5.5	Read
CyberGym	83.1	78.8	81.8	GPT-5.5 is close to Mythos, Opus 4.8 trails
ExploitGym, intended working exploits	157 of 898	NA	120 of 898	Mythos leads, GPT-5.5 is strong
ExploitGym, total flags captured	226	NA	210	Gap narrows when unintended exploit paths count
ExploitGym, kernel exploits	12	NA	22	GPT-5.5 leads on this slice
AISI expert cyber tasks	68.6%	NA	71.4%	GPT-5.5 edges Mythos within uncertainty
AISI “The Last Ones” range, April test	3 of 10	NA	2 of 10	Mythos slightly ahead
AISI later testing, newer Mythos checkpoint	6 of 10 on TLO, 3 of 10 on Cooling Tower	NA	3 of 10 on TLO	Newer Mythos checkpoint reopens gap

Sources: CyberGym, Firefox and OSS-Fuzz figures from Anthropic’s Claude Opus 4.8 system card; Mythos cyber context from Anthropic’s Claude Mythos Preview system card; ExploitGym figures from Berkeley RDI and the associated arXiv paper; AISI figures from the UK AI Security Institute’s GPT-5.5 cyber evaluation and its later autonomous cyber capability update. ExploitGym paired each model with a different agent harness and ran with safety filters disabled, so its results measure model-plus-agent-plus-access conditions rather than raw model capability.