Six models on the reverse-engineering desk
This is a desk map, not a scored bake-off. Vendor numbers below are theirs. Where we have no public RE bench (Grok 4.6), we say so.
Claude Opus 5
The workhorse for “here is an owned sample, take it apart.” Public write-ups of Opus 5 on obfuscated binaries describe the same pattern we see: it does not unpack the protector so much as route around it — emulate, search the workspace, go dynamic. Useful. Incomplete if the finding lives inside the packer stub.
On a reverse-skill session it will pick tools and keep going. Grounding still needs an oracle or a debugger. Fluent wrong offsets are the failure mode.
Claude Fable 5
Mythos-class, built for hours-long coding. Cybersecurity safeguards are not a footnote. Anthropic’s own Fable page: many cyber queries are automatically routed to a less capable Opus (cyber → Opus 4.8). You are billed Fable only if Fable actually answers. A “Fable RE run” that downshifted is an Opus run with extra latency.
On OSWorld 2.0, Anthropic reports Fable 5 / 5.1 scored zero when those safeguards intervened. That is the correct reading of a blocked task, not a coding-bench loss.
Claude Fable 5.1
Same safeguard architecture, cheaper cache reads, better long-horizon coding. Millennium’s public note is the RE-adjacent data point we trust: a one-in-a-million crash that other models missed; Fable 5.1 disassembled a vendor library, matched a core dump, and named the bug. That is binary work framed as debugging — which is often how client RE actually starts.
Use 5.1 instead of 5 for long sessions. Do not assume a malware or packer prompt stays on Fable. If the fallback fires, you wanted Opus 5 (or 4.8) anyway — start there.
GPT-5.6 Sol (ChatGPT / general API)
OpenAI’s flagship as of July 2026. Their cyber benches are the most specific public numbers in this set: ExploitBench 73.5% vs GPT-5.5’s 47.9%; ExploitGym 24.9% under a two-hour cap. They also state GPT-5.6 is better at finding and fixing vulnerabilities than at autonomous end-to-end attacks against hardened targets.
Default Sol ships with system-level cyber screens. OpenAI says those screens also block legitimate defensive work. On their “advanced cybersecurity completion” set of sensitive dual-use prompts, general Sol completed 1.5%. That is a refusal surface, not an intelligence score.
gpt-daybreak-blue
Daybreak Blue is not a different brain. On Bedrock the id is openai.gpt-daybreak-blue-5.6-sol: GPT-5.6 Sol for verified defenders, 1M context, stated uses: vulnerability discovery, detection engineering, incident response. Access is gated (Trusted Access for Cyber).
Blue turns off the extra screens so malware analysis and patch validation can run. It is not Daybreak Red / GPT-5.6-Cyber. On the same sensitive-prompt set, Blue Sol completed 2.0% — still refuses exploit-chain class dual-use, including “pentesting production systems.” For HelixCyber’s offline, authorized samples, Blue is the OpenAI lane. Red is a different program. Details in the refusal note.
Grok 4.6
No public ExploitBench / ExploitGym / OSWorld RE split we can cite. Treat it as a general frontier model with a thinner published cyber-policy surface than Fable or Daybreak. That is not “better at RE.” It means the scope gate cannot live in the vendor. Put it in reverse-skill: no target ACT without granted auth and an offline-sample profile.
Until someone publishes a dated, harness-controlled rerun, we will not rank Grok against Sol on binaries.
What we actually assign
Owned APK / ELF / firmware, analyst in the loop: Opus 5 or Fable 5.1 if the prompt stays on Fable; Daybreak Blue Sol if the org is enrolled and the work is defensive. Grok 4.6 only behind the same harness. Never Fable for a prompt you know will trip cyber fallback unless you want Opus 4.8. Never general Sol if the session is malware triage and you keep hitting screens — that is what Blue is for.
The model is the last choice. Route and scope first.
- Anthropic — Claude Fable (safeguards, fallback to Opus)
- OpenAI — GPT-5.6 (ExploitBench / ExploitGym, Daybreak)
- AWS — Daybreak Blue: GPT-5.6 Sol model card
- Infosecurity — Daybreak Blue vs Red (1.5% / 2.0% / Cyber 95% on OpenAI’s sensitive set)
- Scoring reverse-engineering agents