August Bingo Board · Episode 01

August 2026 Bingo Board: Mid-Month Pulse

Mid-August Board pulse: Qwen3.8-Max, Muse Spark 1.2 / Muse Code, Grok 4.6, and Gemini 3.7 Flash are on the watchlist ungraded. Routing table did not flip.

Chris Watkins 6 min read

Listen in my voice · AI narration (ElevenLabs clone)

Loading audio player…
On this page

Facts sync from the machine layer. Grades stay human. Mid-month is for drift, not a full re-rank unless something earned it. The August drops are on the Board as ungraded / watch. I did not invent letters.

Baseline: July 2026 Bingo Board refresh. Live table: Bingo Board.

What the board looked like after July

  • Coding ceiling: GPT-5.6 Sol (independent SWE-bench Verified lead)
  • Daily driver: Claude Opus 4.8 (package + long-session feel)
  • Computer use: Gemini 3.6 Flash
  • Latency knife: Gemini 3.5 Flash-Lite
  • Editor value: Composer 2.5
  • Provisional July add-ons: Opus 5 / Sonnet 5, Flash Cyber, Grok 4.5
  • Open-weight value lane: Kimi K3, DeepSeek V3, GLM 4.5 (provisional)

What landed since the July refresh

Four models and one harness. Facts synced. No Chris grades.

  • Qwen3.8-Max (Alibaba, Aug 3): 2.4T MoE, 1M context, coding + long-horizon cowork. API via Model Studio / QwenCloud. Open weights promised; HF Qwen/Qwen3.8-2.4T-A95B reported Aug 8. License is bespoke qwen3.8-max, not Apache. Ungraded / watch.
  • Muse Spark 1.2 + Muse Code (Meta, Aug 5): new terminal/CI harness co-trained with the model. Persistent async background agents, git worktree fan-out, append-only event log, skills (/plan, /grill, /goal), muse resume. Beta + Meta Model API. Ungraded / watch.
  • Grok 4.6 (xAI / SpaceXAI + Cursor, Aug 12): long-running agents, visual/interactive work. In Cursor and Grok Build; first-week 2x included usage. API $2/$6 per MTok. Vendor chart ties GPT-5.6 Sol on the AA Intelligence Index (Grok 4.6 = 61, Sol Max = 61, Fable 5 Max = 62). That is a receipt to watch, not a letter. Vendor evals I am citing, not grading: CursorBench 3.2 69.9%, DeepSWE 1.1 65.9%, Terminal-Bench v3.0 26%. Ungraded / watch.
  • Gemini 3.7 Flash (Google, Aug 13): workhorse for coding and agents, 1M context. Intro price $0.75/$3.75 per 1M tokens through 2026-12-31 (then $1.50/$7.50). Gemini API, AI Studio, Vertex, Antigravity; Spark uses it. Ungraded / watch.

Also this window: SWE-Bench ProMax (Aug 10, arXiv:2608.09802), 170 multilingual refactor tasks, avg 11.4 files / 261.6 LOC, best model 41.2%. Yardstick, not a Board letter.

GPT-5.6 SKU split: ChatGPT Sol/Luna August builds are not the same as Codex and ChatGPT Work, which are still on the July Sol/Luna. Do not mix those SKUs when you read a receipt.

What actually moved into mid-August

As of the grades I am willing to defend mid-month, the routing table did not flip. I still route Opus 4.8 daily, Sol as the coding ceiling, and Gemini 3.6 Flash for computer-use until I run the new ones.

Harness controls got louder than Arena screenshots

Builder signal keeps clustering on parallel agent setups (Claude Code, Codex, Cursor Agent, Antigravity, Grok Build, now Muse Code), approval modes, and resumable loops. That matches Living in Color W31 and W32: the fight is who wraps the model and what tools it can touch.

Muse Code is the new control-surface story (event log, worktree fan-out, muse resume). Grok Build’s 4.6 bump is real architecture news for long-running agents. Neither automatically wins the wallet fight against Claude Max or Codex Pro.

Gemini Flash receipts got official, and 3.7 is ungraded

Google’s Gemini 3.6 Flash model card and the July Flash family post (3.6 / Flash-Lite / Flash Cyber) are still the receipts behind the computer-use lane. 83% OSWorld-Verified on a cheap Flash tier still stands. 3.7 Flash is the August workhorse add at intro pricing. I have not run it enough to move 3.6 off the computer-use row.

Provisional rows did not buy a promotion

  • Claude Opus 5 / Sonnet 5: still pilot-with-eyes-open. Independent SWE-class receipts are not thick enough for me to rip 4.8 out of daily loops.
  • Kimi K3 / DeepSeek V3 / GLM 4.5: stay in the BYOK / open-weight column.
  • Qwen3.8-Max: Max-class pressure, bespoke license. Not a daily-driver promotion.
  • Grok 4.5 / 4.6: climbing story continues; harness tax still the structural issue. The AA-index tie is a vendor chart, not a Board letter.

If publish week lands a third-party number that actually moves a letter, patch this section. Until then, refusal to re-rank is the editorial call.

Grades I changed (or refused to change)

  • Changed: none. Live-layer facts synced. No human grade event.
  • Refused: inventing letters for Qwen3.8-Max, Muse Spark 1.2, Grok 4.6, or Gemini 3.7 Flash.
  • Refused: treating the Grok 4.6 vendor AA-index tie with Sol as a coronation.
  • Refused: crowning Opus 5 from launch messaging alone.
  • Refused: treating open-weight leaderboard spikes as Board promotions.
  • Refused: treating Grok Build open-source as a SuperGrok Heavy price fix.

Mid-month refusal is a feature. Hype is not a grade event.

What I am routing this week

JobReach for
Daily multi-turn engineeringOpus 4.8 (pilot Opus 5 on non-critical threads)
Hard correctness-critical codingSol on Codex or Fable 5
Computer useGemini 3.6 Flash (3.7 Flash is watch)
Volume scoutFlash-Lite → 3.6
Security-tilted Flash experimentsGemini 3.5 Flash Cyber
BYOK / local experimentsKimi K3 or DeepSeek V3 via opencode
Editor volumeComposer 2.5
New harness to try, not to crownMuse Code (ungraded)

Watching into month-end

  • Independent SWE / agent receipts for the Claude 5 family
  • Whether I actually run Muse Code, Grok 4.6, 3.7 Flash, and Qwen3.8-Max enough to write letters
  • SWE-Bench ProMax as the refactor yardstick vs saturated Verified screenshots
  • Open-weight coding numbers that deserve letter movement
  • Harness economics (Max vs Codex Pro vs Antigravity quotas vs Grok Heavy)

Month-end note (Aug 28/29): what I would actually route work to after a full month on this table.

Share on X