Alibaba's Qwen4-Max arrives with a 1-million-token window and roughly 60% lower task costs
Alibaba released Qwen4-Max on November 16 with 2.4 trillion1 parameters, native vision and a 1-million-token context window on QwenCloud. Chip stocks fell hard on Monday, so cost questions are getting louder.
Koda editorial team · 17 November 2026 · 6 min read
Iran targeted a U.S. base in Bahrain with drones for the second time in three days and claimed a mine sank a supertanker in the Strait of Hormuz.
Nvidia's hardware and its leadership are telling different stories: Vera Rubin posted the fastest time on six of seven benchmarks in its MLPerf Training v6.1 debut, NVIDIA says, while fallout from Jensen Huang's exit as CEO helped sink the SOX chip index 5.3% and more than 160,000 people have signed a letter asking the board to reinstate him23.
Today: Measure your own costs before you switch models. The steps
Alibaba released Qwen4-Max on November 16 with 2.4 trillion1 total parameters, native vision and a 1-million-token context window, available through QwenCloud API. Alibaba reports roughly 60% lower task costs than Qwen3.8-Max on its own Work360 benchmark, and Digital Today says it beat Claude Fable 5.1 and Kimi K3.1 Pro on selected coding tests.
Why it matters. The 60% saving is measured against Alibaba's own Qwen3.8-Max on a proprietary benchmark1, so it says nothing yet about cost against Claude Fable 5.1 or Kimi K3.1 Pro.
NVIDIA Sweeps MLPerf Training v6.1 as Vera Rubin Debuts
NVIDIA says Vera Rubin posted the fastest time to train on six of seven benchmarks in its first MLPerf Training v6.1 submission.
Why it matters. Buyers planning training clusters now have a benchmark case for choosing Vera Rubin over older NVIDIA chips, which could cut the rental and resale value of existing capacity.
Tesla is taking pre-orders from select private buyers for the $29,990 Cybercab, a two-seat robotaxi with no steering wheel or pedals. It drives unsupervised only inside Tesla's Robotaxi service zones, which cover designated pockets of five Texas cities and three Florida cities.
OpenAI launched ChatGPT Wallet, which lets AI agents make purchases without accessing users' actual card details. Every transaction requires a security check, and the rollout starts with US ChatGPT Pro users.
Chip Stocks Sink 5.3% and Huang Exit Fallout Spreads2
Last week, Jensen Huang's exit as Nvidia CEO cost the Nasdaq just 0.39%.2 On Monday it became a rout: the Nasdaq fell 2.5%, the S&P 500 nearly 2% and the SOX chip index 5.3%.
Why it matters. Chip stocks lost more than twice what the Nasdaq did, a sign that investors now treat Nvidia's leadership change as a risk to the whole sector rather than to one company.
The Nasdaq fell 2.49% in the 16 Nov session, the largest index move, with the S&P 500 down 1.92% and WTI crude down 3.62%; in the 17 Nov session Bitcoin slipped 1.13% and the crypto Fear & Greed index read 12, Extreme Fear.
S&P 500
7,416.59
-1.92%
Nasdaq
24,509.76
-2.49%
Bitcoin
$64,565.31
-1.13%
Ethereum
$1,815.19
-1.11%
Crude Oil (WTI)
$85.35
-3.62%
Crypto Fear & Greed
12 / 100
Extreme Fear
Levels from Yahoo Finance4, 16 Nov and 17 Nov; alternative.me5, 17 Nov.
Go deeper
The Koda Signal, 17 November 2026
Alibaba's Qwen4-Max arrives with a 1-million-token window and roughly 60% lower task costs
TF Code's real bet is identity: each agent run authenticates against a ToothFairyAI workspace, which keeps shared automations tied to a known team context once several people trigger them. Engineering groups already running on ToothFairyAI get a coding agent their internal tools can call through an SDK or REST API, though with zero independent reviews on record, adoption starts with your own trial on a bounded task. Teams outside that platform have little reason to move off Claude Code or Aider yet.
Test the saving on your own prompts before you trust the headline.
Alibaba reports Work360 task cost falling from $4.43 to $1.67, a 62% drop on its own benchmark, and a Monday chip selloff of 5.3% on the SOX index makes budget-conscious builders listen closely2. The mechanism worth acting on sits in your own prompts: trimming oversized ones can save 2.7 times more than the 25% price cut. Run that trim first, then replay your hardest tasks on Qwen4-Max with the 1-million-token window. Only then decide whether the new model earns a place in your stack.
1Replay your hardest tasks. Take a few real jobs from your workload and run them through Qwen4-Max on QwenCloud, logging tokens in and cost out. Alibaba's Work360 figures come from its own benchmark, so your numbers are the ones that count.
2Trim your prompts first. Cut oversized context from the prompts you send most often and compare the bill before and after. Per Alibaba's own figures, that trim can save 2.7 times more than the 25% price cut.
3Try TF Code on a shared automation. Set up one workflow where several teammates trigger the same agent, and check that each run authenticates against your ToothFairyAI workspace. Confirm the team context shows up in the logs before you widen access.
Sources and checks
Every story links to its outlet. Below, by source, are the 14 figures we checked: 0 verified, 7 reported and 7 estimated.
1Digital Today. 2 figures on this page, 17 Nov 2026.ReportedEstimated
2CNBC. 4 figures on this page, 15 Nov 2026.Estimated