Anthropic AI Updates: October 17, 2026

1. Claude Mythos 5 Beats the Best Published Quantum Circuits on Eight Benchmark Primitives

Anthropic. Anthropic gave Claude Mythos 5 plain-language specs for eight quantum circuit primitives (chemistry, optimization, simulation, number theory) plus an automated evaluator that checks correctness and scores qubits multiplied by gates, and let it iterate in a closed loop. Claude found designs using both fewer qubits and fewer gates than the best published circuits in every case. For elliptic-curve point addition, a 53-hour run produced the first reversible divsteps modular inversion with no stored branch history (verified in Lean), including a variant with 12% more qubits but 17x fewer gates. A Berlekamp-Massey based Reed-Solomon decoder cut the costliest step of the DQI optimization algorithm by 12x. The work came out of Anthropic’s STEM Fellows Program with Harvard and QuEra collaborators. Source

2. Claude Recovers How the NIST ECDSA Curve Seeds Were Generated

Anthropic. An internal Anthropic research model worked out that the long-unexplained seeds for NIST P-192, P-224, P-256, P-384 and P-521 were not hashes of a phrase, as rumored. They were output of the 1990s C library lrand48 PRNG seeded with Unix timestamps from May 1997 to April 1998 (P-256 is candidate 1842 of its run). Claude spotted top-bit patterns, recovered the truncated LCG states with lattice reduction, and also explained several base points, though B-163 and the P-224/256/384/521 base points remain unexplained. Anthropic stresses this is not an attack and actually makes deliberate backdoors much less plausible, so no action is needed. Source

3. Anthropic Reports Agent Hours Outnumber Employee Hours 6 to 1 as Junior Technical Hiring Falls

Anthropic. A new economics report using internal employment, telemetry, calendar and hiring data finds that agent work hours passed human hours in March 2026, and by August interactive agent hours outnumbered employee hours 6 to 1. The junior share of technical staff fell from 21% to 14% since early 2025, and junior technical hires fell from 27% to 9%. Output per person rose (more code, more deals closed per rep, more contracts processed), and human attention became the bottleneck: reading all agent output would take non-technical staff half their week. Hiring managers now weight autonomy and judgment over domain expertise, and Anthropic has set up an early warning system for at-risk roles and prioritizes internal redeployment. Source

4. Survey Finds 91% of Anthropic Staff vs. 39% of US Adults Expect Nobel-Level AI Discoveries by 2030

Anthropic. A companion survey compared Anthropic staff with a nationally representative sample of 10,980 US adults (fielded August 2026 by Morning Consult). Both groups largely agree on what AI can do today but split on the pace of progress: staff expect AI to do 75% of their current work by 2030 on average, versus 28% for US workers. Both groups use a similar mental model, though: at the same level of expected automation of their own job, they forecast about the same unemployment rate. The public named customer preference for humans, data privacy, reliability and regulation as the biggest obstacles to workplace adoption. Source (summary)

5. Claude Code Team Details Its Code Review Process as Merged PRs Doubled and Reverts Fell 40%

Anthropic. The Claude Code team published how it reviews code when Claude writes about 90% of the codebase. From June to September, merged PRs more than doubled while the share reverted fell by more than 40%. Most review happens before a PR exists (Claude must show test, build or screenshot output, then a fresh-context subagent reviews the diff), every PR gets an automated review tuned by short CLAUDE.md and code-owner review files, and humans focus on risk-tiered approvals with fast rollback. The target is at most one human in the loop per PR, measured by approvals, comments and wait time, with revert and incident rates as guardrails. The post includes an adoption checklist. Source

6. Claude Code 2.1.302 Adds Auto Mode Classifier Model Override and Flags Standalone Claude Design Shutdown

Anthropic. Claude Code 2.1.302 adds CLAUDE_CODE_AUTO_MODE_MODEL to choose the model used by the client-side auto mode classifier (server-side auto mode is unaffected) and an Idle compaction toggle in /config for the idle compaction feature introduced in 2.1.299. It also adds a notice to /design-sync and /design-login that standalone Claude Design at claude.ai/design closes on December 14, 2026, which teams using those commands should plan around. Cloud session routines now show a banner explaining why the last run did not start, such as hitting an organization’s concurrent session limit. Source