Waymo Pulls Robotaxis Off Freeways Everywhere
Waymo's highway-capable robotaxis were carrying freeway rides; they are now limited to surface streets after California's DMV ordered them off highways for driving into construction zones.
Google's Gemini 4 Neon scored 61 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57. Opus still leads on Terminal-Bench 4.1 and FrontierCode 2.0.
Koda editorial team · 20 November 2026 · 7 min read
Listen · 22 min
Google released Gemini 4 Neon on Nov. 18, priced at $3 per million input tokens and $15 per million output tokens1, with a 1-million-token context window. It scored 61 on the Artificial Analysis Intelligence Index, ahead of Claude Opus 5.5 at 59 and GPT-6.2 Sol at 57, though Opus still leads on Terminal-Bench 4.1 and FrontierCode 2.0.
Why it matters. A third-place Artificial Analysis Intelligence Index score, ahead of Claude Opus 5.5, now costs a midtier $3 per million input tokens1, a price teams picking a default model should check against their current spend.
Go further: today's paper, Storing the Gemma 4 K=V Cache Once: Exact Global-KV Memory Savings in a Paged Serving Engine with Unchanged Attention Kernels (arXiv, 17 Nov).
Waymo's highway-capable robotaxis were carrying freeway rides; they are now limited to surface streets after California's DMV ordered them off highways for driving into construction zones.
OpenAI's ChatGPT for Sales links to Salesforce and HubSpot to research prospects and draft meeting briefs and follow-ups.
Why it matters. Because it reads CRM records directly, the quality of its briefs depends on how clean Salesforce or HubSpot data is, which shifts sales-ops spend toward data hygiene.
Snowflake made Code Bundles generally available on Nov. 19. One EXECUTE CODE BUNDLE command now runs Python, Java or Scala, including Spark and ML Jobs, without building containers.
Why it matters. Move scheduled ML Jobs off custom containers once a Code Bundle runs them from a Snowflake task, which cuts out a build step your platform team maintains.
Helios claims it can read roughly 1,000 blood proteins from a few drops collected with its proprietary device2, a test meant to flag health issues before symptoms appear, with results delivered through an AI health coach. The startup was founded by PhDs Omar Abudayyeh and Jonathan Gootenberg with Yousef Badran, and access is currently waitlisted.
Carta's Q3 2026 US pre-seed data shows median SAFE caps rising across every round size, while total pre-seed dollars grew and deal count fell. Among companies that later raised a seed round, most SAFEs converted at their valuation cap. The pattern describes a funding market tightening in specific, measurable ways: more money concentrated in fewer deals.
OpenAI reports that design changes proposed by its models helped shrink the circuit area of key parts of its Jalapeño AI chip by an estimated 10% to 13%.3 Applied with AMD, the same approach produced up to 14% improvements on AMD's flagship GPU designs, placing model-generated proposals directly in the chip design process.
China had long resisted formal nuclear talks with the US. Now Xi, Trump and Putin have pledged a nuclear-risk dialogue in Shenzhen, with no time frame and no commitments to cut arsenals.
The Nasdaq fell 2.25% on the 19 Nov session, alongside a 2.02% drop in the S&P 500 and a 0.52% slip in WTI crude; separately, the crypto Fear & Greed index reads 12, Extreme Fear.
S&P 500
7,175.83
-2.02%
Nasdaq
23,879.20
-2.25%
Bitcoin
$61,966.18
-1.62%
Ethereum
$1,750.76
-1.35%
Crude Oil (WTI)
$82.83
-0.52%
Crypto Fear & Greed
12 / 100
Extreme Fear
Levels from Yahoo Finance6, 19 Nov and 20 Nov; alternative.me7, 20 Nov.
Icons8's MCP server lets a coding agent pull icons and illustrations from style families drawn by one in-house team, which targets the patchwork look generated screens usually ship with. The agent route sits behind a paid plan, and a deep catalog only helps if the agent chooses well from it; no one has documented how well it does. Product teams generating UI with agents should trial it on a single screen before standardizing.
The cheapest model is the one that is right most often on your work.
Gemini 4 Neon's input token rate runs 50% above GPT-6.2 Sol's, yet per correct result your own success rates decide which model is cheapest. Opus still leads on Terminal-Bench 4.1 and FrontierCode 2.0, so coding-heavy teams should not assume the index leader wins their workload. Pair that test with a tool like the Icons8 MCP server, where a consistent output cuts the rework that inflates cost per finished task. Run your own tasks this week and set your default model from the results.
Below, by source, are the 15 figures we checked: 3 verified, 10 reported and 2 estimated.
Verified confirmed by an independent searchReported as published by the named sourceEstimated approximate, in the source or in our check
Signal 238 · 20 November 2026 · 18 sources · How every number is checked · Corrections
237 editions so far, one a day. Unsubscribe anytime.
Forward this to one operator you work with. Your referral link is in every email; milestones at koda.community/refer.