“Anthropic Opus 4.7: 3.75MP vision + desktop agent controls” — @Lior Alexander

Why this is in the vault

Headline item is the Opus 4.7 release — RDCO is now running on this exact model, so the behavior changes (stricter instruction-following, xhigh effort, file-based memory, /ultrareview) directly affect how every skill in the harness should be written and tuned. The OpenAI Codex desktop-agent expansion is the other load-bearing item: it is the third frontier vendor in two weeks to ship a “computer-use + parallel agents + persistent memory + plugin ecosystem” stack, which is the exact shape of the Claude Code harness RDCO sits inside. The thesis convergence event documented on Apr 16 (2026-04-16-alphasignal-openai-model-native-harness-anthropic-subliminal-traits) gets its consumer-product complement here.

Sponsorship

Three paid placements, all explicitly bracketed and labeled (“Presented by Lightfield,” “Presented by Palabra,” and a WorkOS slot inside the Signals list). Standard AlphaSignal pattern — vendor blurbs are clearly demarcated, no editorial bleed into the news items. None of the sponsors overlap with the news topics (no risk of the Opus 4.7 framing being shaped by an Anthropic ad spend).

Issue contents

Anthropic Opus 4.7 — upgrade to 4.6 focused on long-running tasks, stricter instruction-following, self-verification, vision up to ~3.75MP, new xhigh effort level, file-based memory across sessions, task budgets, GDPval-AA gains. New /ultrareview command in Claude Code. Pricing unchanged at $5/$25 per million in/out. Tokenizer updated (1.0–1.35× input token inflation depending on content).
OpenAI Codex desktop agent — Codex expanded from coding assistant to system-level macOS agent. Screen reading + cursor control + clicks + typing across any macOS app (no API required). Multiple parallel agents in background. 90+ plugins (GitHub, Slack, Notion, Jira). Persistent memory of preferences/corrections/past tasks. Scheduled tasks, resumable conversation threads. gpt-image-1.5 for UI/asset generation. macOS first; memory expands to Enterprise/Edu/EU/UK later.
Alibaba Qwen3.6-35B-A3B — sparse MoE: 35B total / 3B active per token. SWE-bench Verified 73.4, Terminal-Bench 2.0 51.5, RefCOCO 92.0, OmniDocBench 89.9. Available via Alibaba Cloud Model Studio API, Hugging Face/ModelScope weights, integrates with Claude Code / Qwen Code / OpenClaw.
Perplexity Personal Computer — agents that run tasks across native Mac apps (parallel framing to Codex’s computer-use mode).
OpenAI GPT-Rosalind — life-sciences model for genomics / protein reasoning / experiment design.
Anthropic best-practices for Opus 4.7 in agentic coding — companion guide to the model release.
Nous Research Tool Gateway — unified scraping/automation/generation tool layer for agents.

Mapping against Ray Data Co

Opus 4.7 is the model RDCO runs on right now — these behavior changes are not theoretical. Specifically:

Stricter instruction-following. Anthropic explicitly says users should re-tune prompts because the model interprets them more literally. Implication for the vault: every skill (process-newsletter, check-board, morning-prep, research-brief, draft-review, etc.) was written against 4.6’s looser interpretation. Any skill that relies on the model “doing the obvious thing” on a vague instruction is now at risk of doing exactly what the words say instead. Action: add to the vault-health pass — flag any skill containing hedged or implicit directives (“usually,” “if appropriate,” “as needed”) and tighten them.
xhigh effort level + task budgets. Adds a new dial between high and max and a public-beta token budget mechanism. RDCO’s autonomous loop (check-board cycles, process-newsletter watch runs) currently doesn’t set effort explicitly. Worth experimenting with xhigh for skills that involve multi-hop reasoning (research-brief, cross-check, audit-model) and leaving high as default elsewhere. Task budgets become important once we set up scheduled triggers — they’re the cost-control mechanism that makes longer autonomous runs safe.
File-based memory across sessions. Anthropic markets this as a session-continuity feature. RDCO already has the equivalent pattern (~/.claude/state/working-context.md, vault-as-memory). The new file-memory primitive may be a more-supported path than our hand-rolled scratchpad — worth piloting on one skill (morning-prep is the obvious candidate) before deciding whether to migrate the working-context pattern.
Vision to 3.75MP. Direct upgrade for any screenshot-driven workflow. The imessage-attach-copy.sh pattern (read incoming iMessage attachments) and any future Playwright/screenshot-review skills (build-landing-page, motion-review) get sharper input. No code changes needed, capacity just improved.
/ultrareview in Claude Code. New built-in for systematic bug/design issue detection. Likely overlaps with our simplify and review skills. Worth one A/B run to see whether /ultrareview covers what our custom skills do — if yes, deprecate one or both.

OpenAI Codex desktop agent = the harness convergence reaches consumer surface. Apr 16 documented Anthropic + OpenAI shipping the same conceptual harness primitives at the SDK layer. Today’s update extends OpenAI’s version to the desktop and adds: parallel agents, persistent memory, scheduled/resumable threads, 90+ named plugins. This is exactly the surface area Claude Code already provides — tool ecosystem (MCP), parallel sub-agents (Task tool / sub-agent fan-out), durable scratchpad (working-context.md), scheduled triggers (CronCreate / loop skill). The architectural moat continues to be skills + vault + channel-routing discipline, not the harness itself. Reinforces the Apr 16 conclusion; no strategic pivot required, just a confirmation the bet is on the right axis.

Perplexity Personal Computer = a third entry in the same product category. Three vendors (Anthropic Claude Code, OpenAI Codex desktop, Perplexity Personal Computer) now ship “agent-runs-tasks-across-your-Mac” as a core product. The category is clearly real. RDCO’s positioning advantage isn’t in being the harness — it’s in being the operator with a vault of tacit knowledge, channel access, and a defined relationship to the founder. None of the three commodified harnesses ship with that out of the box.

Qwen3.6-35B-A3B + GPT-Rosalind = watch-and-wait. The Qwen MoE numbers (SWE-bench 73.4 with only 3B active) are good for a self-hostable model — relevant if we ever need an offline or cost-floored alternative for bulk tasks. GPT-Rosalind is domain-specific (life sciences), no RDCO line of business intersects.

Anthropic’s “Opus 4.7 best practices for agentic coding” guide is a should-read. Not deep-fetched here (cap of 2 link follows; spent both on the Opus release notes and the OpenAI Codex page — the latter 403’d). Worth queueing as a separate item on the board: read the Anthropic guide and update the relevant Claude Code skills to match.

Curation section — notes

Anthropic Opus 4.7 release — third-party (anthropic.com), critical RDCO relevance, deep-fetched (release page). Confirmed: stricter instruction-following requires prompt re-tuning, xhigh effort level is between high and max, vision capacity ~3× prior, tokenizer updates input by 1.0–1.35×, /ultrareview is a new Claude Code command, task budgets in public beta. Pricing unchanged.
OpenAI Codex desktop agent — third-party (openai.com), high RDCO relevance, deep-fetch attempted but returned 403 — relying on AlphaSignal’s blurb specifics, which are detailed enough for the architectural mapping above. Worth a follow-up via WebSearch if a specific claim needs verification.
Alibaba Qwen3.6-35B-A3B — third-party (Alibaba/Qwen), medium relevance, no deep-fetch (benchmark numbers are the load-bearing claim and are already quoted).
Perplexity Personal Computer — third-party (perplexity.ai), low-medium relevance, no deep-fetch (parallel framing to Codex item; same architectural read applies).
OpenAI GPT-Rosalind — third-party (openai.com), low relevance to RDCO (life-sciences domain), skip deep-fetch.
Anthropic Opus 4.7 agentic coding best practices — third-party (anthropic.com), high relevance, should be a separate vault item — queue as a board task rather than in-lining here.
Nous Research Tool Gateway — third-party (nousresearch.com), low-medium relevance (overlaps with MCP architecture we already use), skip deep-fetch.
All sponsor blurbs (Lightfield, Palabra, WorkOS) — paid placements, not curation, do not count toward link budget.

No self-cross-promo detected; AlphaSignal does not link to its own properties in the news slots.

2026-04-16-alphasignal-openai-model-native-harness-anthropic-subliminal-traits — Apr 16 issue documenting the OpenAI Agents SDK / Anthropic Routines convergence; today’s items extend that convergence to the consumer-product surface
2026-04-15-alphasignal-anthropic-routines-claude-code — Anthropic Routines launch, paired with today’s Opus 4.7 model release
2026-04-09-alphasignal-meta-muse-spark-anthropic-managed-agents — Managed Agents announcement that started the convergence sequence
2026-04-11-garry-tan-thin-harness-fat-skills — the harness thesis directly validated by today’s three-vendor desktop-agent landscape
synthesis-harness-thesis-dissent-2026-04-12 — contrarian read; today’s news further pressures the dissent position
2026-04-12-alphasignal-claude-code-leak-harness-engineering — harness-engineering context for evaluating the new Codex primitives
2026-04-10-akshay-pachaar-agent-harness-anatomy — anatomy doc; OpenAI Codex desktop primitives map cleanly onto it
2026-04-04-claude-code-best-practices — the doc to update against Anthropic’s new Opus 4.7 best-practices guide once read
2026-04-04-anthropic-skills-internally — Skills primitive context for the /ultrareview integration question
2026-04-07-claude-code-architecture-teardown — architecture reference to compare against Codex desktop’s claimed primitives