The quarter the open frontier went permissive — and a flagship got switched off.

Entered the picks

• Qwen 3.6-27B · April 22
  Took the first-pick slot in both chat.top and docs.top, displacing Qwen 3.5 27B. A dense 27B claiming to beat the prior 397B MoE flagship while staying single-GPU.
• Qwen 3.6-35B-A3B · April 16
  The MoE sibling; now the first pick for agents.top. Paired with the dense 27B it defines the 24 GB-tier decision.
• North Mini Code (30B-A3B) · June 5
  Cohere's Apache-2.0 agentic coder entered coding.top and coding.high — the first Western-lab model to earn a coding slot in the 24 GB tier.
• Gemma 4 — relicensed to Apache 2.0 · April 2
  The licence change mattered more than the model. Gemma 3 shipped under custom Gemma Terms; Apache 2.0 unlocked it for commercial work and it now holds slots across chat, docs and image tiers.

Moved

• Llama 4 Scout — demoted out of docs.top · April
  The 10M context number is real for retrieval but not for synthesis; training capped around 256K and quantization compounds the limit. We stopped leading with it and said why.
• RTX 5090 — the price round-trip · April–June
  Floor moved $3,000 (April) → $3,799 (early May) → $2,910 (May 28, ASUS TUF AIBs) → $3,500 (June, as the sub-$3K allocation dried up). We published each move and reversed our own "firming" narrative when the evidence reversed.
• MoE overtook dense as the default shape of a pick · ongoing
  Across the quarter the 3B-active MoE (30B-A3B, 35B-A3B) became the answer at almost every tier above 16 GB — same weight footprint as a dense 30B, 2–3× the generation speed. The calibration table now shows this on a single card.

Left the picks

• GLM-5.1 — dropped from local coding picks · April
  At 744B total it was never locally realistic. It stays as a hosted frontier comparator; pretending it was a local pick was the error.
• Qwen 2.5 Coder, DeepSeek Coder V2 Lite, Phi-3.5, Gemma 2 · April
  The last of the 2024–25 generation left the picks entirely, replaced by Qwen 3.5/3.6-era models. Every competitor we surveyed still recommends at least one of these.
• Mac mini at $599 — discontinued · May 1
  Apple removed the base config on May 1; the entry Mac floor moved to $799/512GB. The cheapest credible Mac for local AI got $200 more expensive without changing.

Hardware

• Library expanded 11 → 23 picks · April 20
  Added the used-market tier (RTX 3090, RTX 4090, A6000 48 GB), the Blackwell middle (5080, 5070 Ti, 5060 Ti), the Mac ladder (mini M4 Pro, MBA M5, M5 Pro MBP, Mac Studio M3 Ultra 96 GB), unified-memory boxes (Framework Desktop Strix Halo, DGX Spark) and Dual 5090.
• Apple raised Mac prices across the line — June 25 · June 25
  DRAM/NAND shortage. Mac Studio M3 Ultra 96 GB $3,999 → $5,299 (+$1,300); M4 Max 64 GB $3,199 → $3,799; mini M4 Pro $1,399 → $1,599. We recomputed every Mac break-even verdict — the M3 Ultra's payback vs a $100/mo plan went from ~36 to ~56 months.
• The DRAM shortage became structural, not a blip · ongoing
  Micron guidance points past 2028. It props up every floor in the table — the RTX 3060 came back into production in June and still could not get back to its 2021 launch price.

Cloud pricing

• Claude Fable 5 — launched, then switched off · June 9–12
  Released June 9 at $10/$50, suspended June 12 under a US export-control directive, and still suspended when the quarter closed. Nineteen days in which no amount of subscription money could turn the best hosted model back on. It was restored July 1 — in Q3.
• Claude Opus 4.7 → 4.8, price held flat · May 28
  Standard pricing stayed $5/$25 — no tokenizer surprise this time — while fast mode got 3× cheaper ($10/$50 vs $30/$150).
• The cheap-tier war opened · May–June
  Haiku 4.5 ($1/$5), Gemini 3.5 Flash ($1.50/$9), Grok 4.3 ($1.25/$2.50) and Sonnet 5 (intro $2/$10) all landed in the calculator. The floor for credible hosted inference fell hard — which stretches local payback for light users and changes nothing for heavy ones.
• GPT-5 pricing was wrong on our own site, and we fixed it · May 28
  We had published $0.625/$5 against a real $1.25/$10 — a pre-existing data error, not drift. Corrected May 28 and the calculator assertion updated with it.

Methodology + evidence

• "What we measured" → "What the community has measured" · May 16
  The calibration page claimed we ran the benchmarks. We cross-verified published community figures instead. The claim was reframed to match reality — the single most important honesty fix of the quarter.
• Every page became a real page · May 3
  Build-time prerender shipped: every URL now serves its own HTML with its own title, description and canonical. Before this, every page on the site returned byte-identical markup claiming to be the homepage.
• The site became machine-readable on purpose · May–June
  llms.txt + llms-full.txt, per-bot allow rules, Content Signals, an RFC 9727 API catalog, and Markdown content negotiation — agents can fetch any page as clean markdown. Built in-code because the vendor toggle costs $240/yr and Path 2 says zero infra cost.
• Memory-fit math replaced tier-overlap guessing · April 29
  Fit badges now compute weights + KV cache at your actual context depth, minus the memory your other apps are using. Same hardware, different context, different verdict — because that is how it really works.

What we got wrong

A currency claim is only worth anything if the corrections are published too. These were live on this site during Q2 2026 and are fixed now.

• We told AMD users to avoid ROCm for months after the bugs were fixed
  The MoE-on-ROCm crash story was true when written. llama.cpp #20024 was fixed on both backends on March 12 and #19880 closed February 28 — we were still publishing "still open" in July. Corrected July 17.
• We called Jan AGPLv3 and reasoned from it
  Jan relicensed to Apache-2.0 in May 2025 — eleven months before we wrote the guide. The advice ("pick Jan if your shop blocks copyleft") was backwards for the whole time it was published. Corrected July 17.
• We cited an estimate engine as a benchmark
  Two planner why-strings quoted WillItRunAI tok/s figures as measurements. Its own footer says the numbers are calculated from specifications. Removed July 17; the lesson that caused it is retracted in MILESTONES.

Next step

See what your hardware runs today→