Living in Color August · Episode 01

Living in Color W31: What Actually Moved

August kickoff for AI in Living Color: Opus still daily, Sol as coding ceiling, Gemini Flash tiers eating computer-use, open-weight pressure rising, and why harness controls matter more than the logo on the model.

Chris Watkins 4 min read

Listen in my voice · AI narration (ElevenLabs clone)

Loading audio player…
On this page

August opens with a noisy timeline and a quieter truth: the board did not flip upside down. A few lanes got sharper. Here is what I am actually carrying into the month, grounded in the July Bingo Board refresh plus late-July through Aug 2 signal, not the loudest quote-tweet.

Three things that mattered

1. Claude Opus is still the daily driver (ignore the rename churn)

The discourse wanted a coronation fight. My routing table did not. Opus 4.8 still wins the package I pay for every day: multi-turn engineering feel, SWE-bench strength, and sessions that stay coherent when the ticket gets messy.

Opus 5 headlines keep circulating. Daily driver is a habit with receipts, not a press release. If a newer Opus tier shows up with public, graded receipts on the Board, I will re-score it. Until then I am not rewriting my week around a launch thread.

2. GPT-5.6 Sol took the coding ceiling lane

Independent SWE-bench Verified at 96.2% put Sol ahead of Fable 5 on the raw coding receipt. That matters when the ticket is hard and the harness is terminal-native (Codex Pro).

It does not make Sol my aggregate top. Fable still wins overall depth. Opus still wins the long session. Sol wins “highest verified coding number right now.” Keep GPT-5.5 as the proven OpenAI workhorse. Open Sol when correctness is the whole game.

3. Gemini Flash tiers ate computer-use and latency

Gemini 3.6 Flash at 83% OSWorld-Verified as a cheap Flash model is still the real Google story. The official model card and the July Flash family post (3.6 / Flash-Lite / Flash Cyber) are the receipts. Antigravity-style loops get 3.6 as the default brain for me.

Gemini 3.5 Flash-Lite stays the throughput knife (Google’s claim around 350 tok/s). Scout with Lite, close with 3.6. Betting production loops on unconfirmed Pro SKUs is still not the move.

One thing the discourse got wrong

People keep ranking models when the fight moved to harnesses and controls.

Aug 1 to 2 signal is full of the same plot: builders running Codex, Claude Code, Cursor, Grok Build, and friends through real workflows with no single winner; people combining remote Codex + Claude Code cloud agents + Cursor async on purpose; MCP installs that suddenly let an IDE agent touch chain or docs. That is the week. Who wraps the model, what tools it can touch, and whether a human can stop it mid-loop.

If you only watch Arena screenshots, you will miss it.

Board / stack implication

JobReach for (into August)
Daily multi-turn engineeringOpus 4.8
Hard correctness-critical codingFable 5 or Sol on Codex
Computer use / cheap desktop agentsGemini 3.6 Flash
Latency and volume scoutingFlash-Lite, escalate to 3.6
Editor value laneComposer 2.5 inside Cursor
BYOK / open-weight experimentsKimi K3 or DeepSeek V3 (provisional)

Open-weight pressure is louder into August. Kimi K3 weights and DeepSeek V4-Flash discourse are real. On my Board, Kimi K3 and DeepSeek V3 stay in the value / BYOK lane as provisional grades until independent coding receipts settle the way Sol and Opus did. GLM 5.5 rumors are watchlist, not a grade event. Board-sync also flagged thin Google rows (gemma-4-12b, gemini-3-5) with empty benchmarks. Those wait for an editorial pass. I do not invent letters.

Full live table: Bingo Board.

One builder action for the week

Pick your default harness for August and write its tool allowlist down. Model logo second. Permissions first.

That is also the bridge into this month’s authority series on Kelpie: Field Notes Part 1 lands Wed Aug 5 on why spreadsheets fail agent inventories.

Watch next

  • Mid-month Board pulse (Aug 14): what actually drifted after these July lanes
  • Big Mama Production Diary (Aug 7): latency budget that does not kill trust
  • W33 beginner deep dive: what a harness is, in plain language

Signal for this kickoff includes harness bake-offs (Codex / Claude Code / Cursor / Grok Build), Gemini 3.6 Flash model-card receipts, Kimi K3 weights, DeepSeek V4-Flash price-performance noise, and multi-harness builder setups. I cut anything I could not square with Board grades I already signed.

Companion vlog: Living in Color: Week of Aug 3 (8 to 12 min talking head). X/Shorts pack: docs/operating-loop/distribution/august-2026/w31-living-in-color.md.