"Anthropic Claude Sonnet 5: flagship reasoning at 40% of Opus cost" — AlphaSignal

Why this is in the vault

Three Anthropic-native developments in one issue that directly affect the RDCO agent stack: Sonnet 5 changes the cost calculus for agentic workloads (near-Opus quality at 40% of the price), Fable 5's return after government-enforced review is a live data point on AI governance velocity, and Claude Managed Agents just shipped five features that reduce operational pain for production agent deployments. Also: a CMU paper questioning whether most "AI agents" are just scaffolding, which maps directly to the fat-skills/thin-harness framing.

⚠️ Sponsorship

Three paid placements in this issue:

• Datadog ("Build Trustworthy AI Agents, Faster") — Datadog Agent Observability product pitch embedded between news items
• Render ("Stop wiring up queues, workers, and retries") — Render Workflows with promo code RENDER-ALPHASIGNAL for $50 credits
• Datalab ("Chandra OCR 2 tops olmOCR") — presented-by placement inside the Signals section; OCR benchmark claims are sponsor-provided

Editorial coverage of Anthropic items appears organic. Sponsor claims (Chandra OCR benchmarks, Render product comparisons) should not be taken as independent analysis.

Curation section

Top News 1 — Fable 5 returns after US government security review

• Amazon researchers found a jailbreak that prompted Fable 5 to identify software vulnerabilities; government pulled the model for ~3 weeks
• Anthropic trained a new safety classifier targeting the specific bypass; blocks the exploit in >99% of cases
• Tradeoff: classifier occasionally flags benign coding/debugging requests; those sessions fall back to Opus 4.8 with a visible notice
• Live now on Claude.ai, Claude Code, and Cowork; paid plans get Fable 5 for up to 50% of weekly usage limits, then usage credits
• /feedback in Claude Code routes false-positive reports back to Anthropic for classifier tuning

Top News 2 — Claude Sonnet 5 ships

• Near-flagship agentic performance: 63.2% vs Opus 4.8's 69.2% on agentic coding benchmarks
• Pricing: $2 input / $10 output per million tokens (intro, through Aug 31), then $3/$15
• That's roughly 40% of Opus standard pricing for near-equivalent agentic capability
• 1M token context window, 128K token output limit
• Default model for Free and Pro plans; API access via claude-sonnet-5
• Self-checks its own work without being asked; completes complex tasks where prior Sonnets stalled

Top Repo — Claude Managed Agents: 5 new features

• Agent Overrides: start a session from a saved agent config and swap model/system prompt/tools for just that session — no agent duplication for small variations
• Streaming session event deltas: real-time updates as agent works, enabling reactive UI instead of polling
• New webhook event types: server-side notifications on session state change; eliminates fragile long-lived HTTP connections
• Credential injection scoping: per-user API token isolation for multi-tenant agent deployments
• Observability tab in Console: per-session token usage and tool call logs for debugging
• Onboarding: /claude-api managed-agents-onboard in Claude Code, or claude-cookbooks on GitHub

Signals

1. NVIDIA splits a 30B model into two cooperative pieces to generate text 2.42x faster — inference efficiency without a smaller model
2. (Datalab-sponsored) Chandra OCR 2: 92.1% on tables, 93.7% on tiny text vs olmOCR benchmarks
3. Allen AI open-source toolkit converts messy PDFs into clean LLM training data — 18K GitHub stars
4. CMU paper: most "AI agents" are just scaffolding, not true agents — 689 likes
5. Jackrong 35B open-source multimodal coding model with tool use and long context — 12.6K downloads
6. Open-source Rust browser as a single binary, loads pages in 85ms — positioned for web scraping and AI agent use

Mapping against Ray Data Co

Sonnet 5 pricing decision: At $2/$10 intro vs $3/$15 post-Aug-31, this is a direct input to the RDCO model-selection layer. For the always-on COO agent running long agentic sessions, switching from Opus to Sonnet 5 could cut per-session cost ~60% through August while keeping agentic benchmark quality within 6 percentage points. Worth tracking actual task-completion parity over the next few weeks before committing to a default swap.

Managed Agents features: Agent Overrides and credential injection scoping are the two that matter most for RDCO's multi-session architecture. Overrides would let the channels agent swap its system prompt per conversation type without maintaining separate agent configs. Credential scoping is relevant if RDCO ever runs agents on behalf of multiple phData client contexts.

CMU scaffolding paper (Signal 4): Maps directly to the fat-skills/thin-harness thesis. If most "agents" are just scaffolding, the competitive moat is in what the skills do, not in the orchestration layer. Worth pulling the paper.

Allen AI PDF toolkit (Signal 3): Potentially useful for the vault ingestion pipeline — converting source-PDF materials (research papers, contracts) into clean markdown for QMD indexing. 18K stars suggests real adoption.

Fable 5 + government review: The 3-week government-enforced pause is a precedent event. Export-control mechanisms now apply to deployed model weights, not just hardware. This is the first documented instance of a live model being pulled for national-security review and returned with classifier modifications — worth flagging for any RDCO risk thinking about model availability SLAs.

Related

• [[2026-07-01-moonshots-sonnet-5-drops-fable-5-ep268]]
• [[2026-06-30-alphasignal-brain2qwerty-meta-bci-open-source]]
• [[2026-07-01-every-fable-5-prompt-library]]