"Vibe Check: Opus 4.8 — Anthropic Should've Rounded Up to 5" — @danshipper / @katie.parrott12
Why this is in the vault
External, same-day hands-on verdict on Opus 4.8 from Every's recurring "Vibe Check" model-eval column — landing the exact day RDCO's COO agent ("Ray") upgraded to Opus 4.8. Direct corroboration of that just-made decision: Every rates 4.8 their most complete model to date, best-in-class on both senior-engineer coding and writing, with one operational caveat (effort-setting sensitivity) that matters for an always-on cron agent.
⚠️ Sponsorship
- Third-party paid sponsor: Lightfield — an outbound/CRM "AI sales agents" product. Standard Every paid ad block ("Want to sponsor Every? Click here"). Unrelated to the model eval; no bias into the Opus verdict. UTM-tagged (
utm_campaign=every). - Every self-promo (note, not third-party): the subscription-upsell pane plugs Every's own AI products — Sparkle, Cora, Spiral, Monologue — and the "Work with us" AI-consulting CTA. Every builds and sells its own AI tooling, so its model evals are not disinterested: a verdict favoring "the model over the app" (i.e., reach for the raw model via your own harness rather than the lab's first-party app) is directionally consistent with Every's business of wrapping models in its own products. The benchmark numbers are still the useful signal; treat the app-vs-model framing as house-aligned.
- Disclosure quality: explicit for the paid ad, structurally visible for the self-promo. No undisclosed relationship detected.
The core argument
Every's framing: "Anthropic is back." After a rough patch (Opus 4.7 "hard to love," OpenAI's Codex desktop app pulling even devoted Claude users to GPT), Opus 4.8 is the first Anthropic release in a year they'd reach for across coding, prose, and everyday work. Four tested findings:
- Best on senior-engineer coding. At extra-high effort, Opus 4.8 scored 63 on Every's Senior Engineer Benchmark, vs 62 for GPT-5.5 and 33.5 for Opus 4.7. Big caveat: "the score drops significantly" at lower effort settings.
- Strongest writing model they've tested. Opus 4.8 at high effort scored 79.6, ahead of Sonnet 4.6 (74.5), GPT-5.5 (73), and Opus 4.7 (63) — with "fewer AI tells than any non-Claude model."
- Best one-shot PowerPoint they've seen on their Every Consulting Benchmark — a well-designed deck telling a clear story, which most models still can't do.
- The model is stronger than the app. 4.8 is good enough to "want to live in Claude," but the split across Chat/Code/Cowork keeps Codex the better daily harness. The verdict is pro-model, app-skeptical.
Net: a near-clean win on capability (the title jokes Anthropic should've called it 5), undercut only by Anthropic's product surface — not the model.
Mapping against Ray Data Co
Unusually load-bearing today: Ray upgraded to Opus 4.8 this morning (2026-05-28). Every's independent eval is same-day external validation of that call.
- (a) 4.8 vs 4.7, specifically. The jump is not incremental on the dimensions RDCO cares about. Senior-eng coding 33.5 → 63 (nearly 2x) and writing 63 → 79.6. These are precisely the two surfaces Ray runs hardest: build/code work and content drafting. Every independently confirms 4.7 was the weak point — matching RDCO's own read in [[2026-05-14-every-opus-4-7-reels-us-back-in]] and [[2026-04-17-every-vibe-check-opus-4-7]] — and that 4.8 fixes it. The upgrade is corroborated, not just vibes.
- (b) Writing surfaces. The writing-test result (top score, fewest AI tells of any non-Claude model) supports using 4.8 for RDCO content drafting — Sanity Check drafts, remix output, vault prose. The "fewer AI tells" finding is directly relevant to the founder's no-slop / voice-match discipline; 4.8 should reduce the em-dash/AI-tell cleanup load. Worth pairing with [[feedback_no_em_dashes]] and the voice-match workflow.
- (c) Operational caveat for an always-on cron agent. The actionable line: top coding scores are at extra-high / high effort, and "the score drops significantly" at lower effort. Ray runs autonomously on cron, where effort/thinking-budget settings are easy to leave at defaults. If 4.8's headline gains are effort-gated, a cron loop running at low effort may not realize the 4.7→4.8 jump on hard coding tasks — and could even underperform expectations while paying Opus prices. This is a config tripwire worth verifying on the agent's actual run settings, not a reason to roll back.
- App-vs-model framing is moot for RDCO. Every's main knock (Chat/Code/Cowork app split) doesn't apply to us — RDCO runs the model through its own Claude Code harness, exactly the "reach for the model, bring your own harness" posture Every endorses. The app critique is irrelevant; the model verdict is what transfers.
Related
- [[2026-05-14-every-opus-4-7-reels-us-back-in]] — prior Every take on 4.7; the "rough patch" this issue says 4.8 ends
- [[2026-04-17-every-vibe-check-opus-4-7]] — the direct predecessor Vibe Check (4.7), same benchmark frame for apples-to-apples comparison
- [[2026-04-17-alphasignal-opus-4-7-codex-desktop-control]] — corroborates the "Codex pulled users from Claude" backdrop
- [[feedback_no_em_dashes]] — "fewer AI tells" finding maps to RDCO's AI-tell / voice discipline
Body rendered cleanly from the Gmail HTML (no web reconstruction needed). Source quotes ≤15 words, paraphrased per copyright discipline. Full eval (Reach Test ratings, pricing, screenshots) is behind Every's site link; this note covers the email's summary findings.