AI News Sep 3 2026
Meta has released four versions of the same model in five months, and the fourth one is where the strategy finally becomes obvious. Muse Spark 1.3 landed this week at exactly the same API price as Spark 1.2, $1.25 per million input tokens, $0.15 per million cached input, and $4.25 per million output, with a 1 million token context. What changed is efficiency: roughly 20 percent fewer tool calls and 25 percent fewer tokens to finish the same work. Put that through Artificial Analysis and Spark 1.3 (xhigh) costs $0.55 per Intelligence Index task while GPT-5.6 Sol (max) costs $0.95 and Grok 4.6 (high) costs $0.94. Those two carry a premium of more than 70 percent for scores that are not 70 percent better.
Google spent the same 48 hours splitting its workhorse tier in two. Gemini 3.8 Flash shipped at frontier-adjacent performance for $0.75 and $3.75 per million tokens, and Gemini 3.8 Flash Cyber shipped alongside it, taking the CyberGym crown at 86.2 percent and finding a critical vulnerability for Google's own Cloud Vulnerability Research team in under two hours, work that normally takes months. OpenAI wired ChatGPT Health into Epic, reaching records for 325 million patients. Here are the 16 stories that matter for September 3, 2026, all from the last 48 hours. For running coverage, bookmark our AI industry news and trends hub.
1. Meta Muse Spark 1.3: Pricing, Benchmarks and What Changed
Meta released Muse Spark 1.3 this week, its fourth Muse Spark release in five months, available through Muse Code and the Meta Model API. Pricing is unchanged from Spark 1.2 at $1.25 per million input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens, with a 1 million token context window. Meta reports roughly 20 percent fewer tool calls and 25 percent fewer tokens than Spark 1.2 to complete equivalent work, and Muse Spark 1.3 (max) scores 62 on the Artificial Analysis Intelligence Index.
Holding price flat while cutting token consumption is a price cut that does not appear in the price list, and it is the more honest kind. Twenty percent fewer tool calls matters specifically for agentic workloads, where each call carries latency and a round trip as well as tokens, so the saving compounds beyond the raw token count. Meta describes Spark 1.3 as handling multiple workloads inside a single conversation, noticing gaps in its own plan, and preserving instructions as tasks, which is the vocabulary of long-running agent work rather than chat.
My take: shipping four versions of one model in five months looks like indecision from outside and reads as discipline from the numbers. Each release has moved efficiency rather than headline capability, which is a harder thing to market and a more useful thing to buy. One caveat on the figures: I have seen output pricing reported as both $4.25 and $4.75 across sources this week, so confirm against Meta's own API page before you build a budget on it. Our AI coding tools hub tracks what teams are running.
2. Muse Spark 1.3 vs GPT-5.6 Sol: Cost Per Task Compared
Muse Spark 1.3 (xhigh) costs $0.55 per Artificial Analysis Intelligence Index task, against $0.95 for GPT-5.6 Sol (max) and $0.94 for Grok 4.6 (high). That places both rivals at a premium of more than 70 percent for the same evaluation suite. Spark 1.3 (max) scores 62 on the Intelligence Index, against Claude Opus 5 at 63.
Cost per task is a far better metric than cost per token and almost nobody publishes it, which is why this comparison is worth sitting with. Per-token pricing tells you what a unit costs, not how many units a model needs, and a cheaper model that takes three attempts is not cheaper. Artificial Analysis computes cost per task by running the same evaluation across models and dividing total spend by tasks completed, which folds verbosity, retries, and reasoning overhead into a single number. On that basis Meta is delivering within one index point of Anthropic's leader at roughly half the cost of OpenAI's.
My take: this is the number I would put in front of anyone still choosing models on a price-per-million-tokens table. Run the same comparison on your own workload before you switch, because the Intelligence Index task mix will not match yours and the ranking can flip. But the general lesson holds, and it is the same one NVIDIA's NOOA framework demonstrated on SWE-bench Verified last month: efficiency of completion beats headline capability for almost every production budget.
3. Gemini 3.8 Flash Cyber: CWE-Bench, Patching and Real Vulnerabilities
Google launched Gemini 3.8 Flash Cyber on September 2, 2026, a security-focused variant of its workhorse model built for vulnerability detection and automated patching. It scores 47.2 percent pass@1 on the CWE-Bench patching benchmark, against 47.8 percent for a leading frontier model, at what Google says is materially lower cost. On an internal benchmark spanning 20 programming languages it exceeds a 70 percent success rate. Google's Cloud Vulnerability Research team used it to identify a critical vulnerability in under two hours, a task that normally takes months, and its Chrome Security team reported the model producing 2.6 times more correct patches to Chrome vulnerabilities than comparable commercial models.
The 2.6x correct-patch figure is the one that should get attention, because patching is where security AI usually falls apart. Finding a vulnerability is a search problem that models handle well. Writing a fix that actually closes the hole without breaking the surrounding code is a reasoning problem, and most tools produce patches that look right and fail review. A model matching a frontier system on CWE-Bench at lower cost, then outperforming commercial rivals 2.6 to 1 on real Chrome patches, is a genuinely different claim from a benchmark score.
My take: the under-two-hours result deserves the same scepticism as any vendor case study, since we do not know how the target was selected or how many attempts preceded it. That said, Google is one of the few organisations that can validate this on its own production codebase and report the result, and Chrome is not a toy. If the 2.6x holds across other codebases, this changes the economics of vulnerability remediation more than it changes discovery.
4. CyberGym Leaderboard 2026: Gemini 3.8 Flash Cyber Takes the Lead
Gemini 3.8 Flash Cyber now leads the CyberGym benchmark at 86.2 percent, ahead of GPT-5.5-Cyber at 85.6 percent, Claude Mythos 5 at 83.8 percent, GPT-5.6 Sol at 83.6 percent, and Gemini 3.5 Flash Cyber at 77.5 percent. That displaces Z.ai's GLM-5.3, which had led CyberGym at 84.5 percent since its August 14 launch.
The composition of that leaderboard is the story. Three of the top five entries are purpose-built cyber variants rather than general models, which was not true two months ago, and the gap between the specialised and general versions of the same family is wide. Gemini 3.8 Flash Cyber at 86.2 percent against Gemini 3.5 Flash Cyber at 77.5 percent is nearly nine points in one generation. GLM-5.3 remains the only model in that group with published open weights, which is the uncomfortable detail: the leader is gated, the runner-up is gated, and the model anyone can download sits third.
My take: a leaderboard where the top entries are all access-controlled and the open entry is a few points behind is roughly the safety outcome the industry said it wanted, and it is fragile. GLM-5.3's weights are already public and cannot be recalled, so the floor for offensive capability is now set by whatever the best open model can do, not by whatever the best model can do. Every point the gated leaders add is a point that does not reduce the floor. Detail on GLM-5.3's release sits in our August 30 roundup.
5. Google's Fairwind Program: Who Gets Access to Flash Cyber
Gemini 3.8 Flash Cyber is available exclusively through Google's new Fairwind Program, which targets government authorities, critical infrastructure operators, and software maintainers. It is not available through the general Gemini API.
Restricting an offensive-capable model to defenders is the same architecture OpenAI applied to Astra two days earlier, gating advanced cyber capability behind partner vetting after declaring it had crossed the Critical threshold. Including software maintainers alongside governments and infrastructure operators is the notable choice, because maintainers of widely used open-source projects are exactly the people who need vulnerability discovery most and have the least budget for it. Whether the vetting is fast enough to be useful to a volunteer maintainer is the practical question.
My take: three labs converged on capability gating within a single week, OpenAI with Astra, Anthropic with the Mythos tier, and now Google with Fairwind. That consensus arrived without regulation, which is worth acknowledging. My reservation is the same as it was on Tuesday: gating works while the frontier is narrow, and CrowdStrike shipped an explicitly offensive model commercially in the same period. The industry is running two contradictory strategies at once and calling both responsible.
6. Gemini 3.8 Flash Pricing and Performance Confirmed
Google confirmed Gemini 3.8 Flash pricing at $0.75 per million input tokens and $3.75 per million output through December 31, 2026, matching the rate Gemini 3.7 Flash carried. It rolled out three weeks after the previous release, continuing the fastest workhorse cadence in the industry, and Google describes it as delivering frontier-level performance at a fraction of the price with reduced verbosity.
Inheriting the December 31 expiry is the detail I flagged as unknown yesterday and it is now answered. Anyone building a 2027 budget on Flash pricing should model $1.50 and $7.50, not $0.75 and $3.75, because the rate doubles on January 1. Reduced verbosity is a real cost lever at these prices given output is billed at five times input, so a model that says the same thing in fewer words cuts bills without any code change. Three weeks between workhorse releases while Gemini 3.5 Pro remains unshipped tells you where Google's engineering effort sits.
My take: Google is winning the tier that generates revenue and losing the tier that generates headlines, and under new leadership from Koray Kavukcuoglu that trade is now a deliberate choice rather than an accident. I think it is the right commercial call and a poor perception call, and perception is what enterprise buyers use when they have not run their own evaluation. Full pricing comparison sits further down this post.
7. Google Pics: A New Image Creation Tool
Google released Google Pics alongside the Gemini 3.8 Flash launches, a new image creation tool. It arrives into a generative image market where OpenAI's gpt-image-2 leads the Arena boards and xAI's Grok Imagine Image 2.0 sits second at 1,320 points in text-to-image and 1,439 in image editing.
Launching a standalone image product rather than folding the capability into Gemini is the interesting decision, and it mirrors what Google did with Gemini 3.5 Transcribe as a separate speech endpoint last week. The pattern suggests Google now treats image, speech, and text as distinct products with distinct pricing and distinct users, rather than as modes of one multimodal model. That is the opposite of where MiniMax and Alibaba are heading with omni-modal single models like H3 and Qwen3.8-Max.
My take: two architectural philosophies are now visible and both have merit. Separate specialised products are easier to price, optimise, and improve independently. Single omni-modal models are easier to build with and cheaper to serve. I suspect the enterprise market prefers the first and the developer market prefers the second, which means both survive. Watch whether Google Pics gets an API, because without one it is a consumer product rather than infrastructure.
8. Best Chinese AI Model September 2026: Kimi K3 Leads at 79.9
Kimi K3 from Moonshot AI leads the September 2026 ranking of Chinese AI models with a composite score of 79.9, ahead of Alibaba's Qwen3.8-Max at 78.6 and Tencent's Hy4 preview at 78.3. The scoring methodology weights reasoning at 17 percent, multimodal at 12 percent, and knowledge at 12 percent across multiple evaluation benchmarks.
A 1.6-point spread across the top three is close enough that the ranking order depends heavily on the weighting scheme, and a different methodology would likely reorder them. What the numbers establish is parity: three separate Chinese labs, using different architectures at wildly different scales, arrived at effectively the same capability level. Kimi K3 runs 2.8 trillion parameters activating 16 of 896 experts, Qwen3.8-Max runs 2.4 trillion with roughly 95 billion active, and Hy4 preview runs 770 billion with 49 billion active. The smallest of the three is within 1.6 points of the largest.
My take: the parameter spread is the finding here, not the ranking. Hy4 preview delivers 78.3 at less than a third of Kimi K3's parameter count, which says architecture and training now matter more than scale in this band. For deployment that is decisive, because Hy4 also ships under Apache 2.0 while Kimi K3 carries a custom licence. Full detail sits in our Kimi K3 review.
9. Claude Fable 5 Tops the Arena Text Leaderboard at 1508.6
Claude Fable 5 ranks first on the Arena text leaderboard overall with a score of 1508.6 and leads six of seven subcategories, making it the highest-rated model for conversation. Separately, Claude Fable 5.1, released September 1, is recorded as the best-ranked model released during September 2026.
Arena rankings measure blind human preference in head-to-head comparisons rather than task completion, which makes them the closest thing to a measure of whether people actually like using a model. That is a different question from capability, and the two diverge regularly. Fable 5 leading conversation while Claude Opus 5 tops the Artificial Analysis Intelligence Index at 63, and Claude Opus 4.7 leads SWE-bench Verified at 87.6 percent, means Anthropic holds three separate crowns with three different models, none of which is its newest release.
My take: holding the top spot on human preference while Fable 5 plateaued at 11 percent of Anthropic's customer spending is the most instructive contradiction in AI right now. People prefer it and companies do not buy it, because preference is measured per conversation and procurement is measured per invoice. That gap explains most of Anthropic's product strategy this year, including shipping Fable 5.1 with a 75 percent cache read cut rather than a capability jump.
10. The September Fall Rush: Four Labs, Four Strategies
OpenAI, Anthropic, Google, and Meta all shipped new models or major updates within roughly 48 hours this week, and each picked a different axis to compete on. OpenAI pushed autonomous agent capability with Astra, which crossed its Critical cybersecurity threshold. Anthropic emphasised price-performance, cutting Claude Fable 5.1 cache reads 75 percent for savings up to 45 percent on agentic workloads. Google focused on coding and security with Gemini 3.8 Flash and Flash Cyber. Meta targeted efficiency, cutting tool calls 20 percent and tokens 25 percent in Muse Spark 1.3.
Four labs releasing in one week is coordination by market pressure rather than by agreement, and the differentiation is the genuinely new part. Eighteen months ago every release claimed the same thing, which was general capability improvement measured on the same benchmarks. Now each lab is claiming a distinct axis, and the axes barely overlap. That is what a maturing market looks like, where competing head-on becomes less profitable than owning a segment.
My take: specialisation is good news for buyers and bad news for anyone maintaining a single-model architecture. If the best agent model, the cheapest agentic model, the best security model, and the most efficient model are four different products from four different vendors, then routing between them is no longer an optimisation, it is the correct architecture. Our AI agent frameworks hub covers the tooling that makes that practical.
11. ChatGPT Health Connects to Epic Records for 325 Million Patients
OpenAI launched an Epic electronic health record integration for ChatGPT Health on September 1, 2026, giving authorised clinicians read-only access to records covering more than 325 million patients. Clinicians can pull appointment notes, laboratory results, medications, and specialist documentation into ChatGPT, or work directly within Epic workflows for pre-visit review and clinical timeline building without leaving a patient chart. UCSF Health is a pilot partner. A separate Healthcare Public Data plugin provides structured access to PubMed, DailyMed, and CMS Coverage.
Read-only is the design decision that makes this deployable. The model can read a chart and cannot write to it, which removes the entire category of risk where an AI system alters a medical record, and it means the clinician remains the only party who commits anything. Epic holds records for a very large share of US patients, so an integration at this level reaches further than any pilot programme could. The public data plugin matters as much, because a model reasoning over a patient chart without current drug labelling and coverage rules is reasoning with half the picture.
My take: this is the most significant healthcare AI deployment of the year and read-only is why it will survive contact with hospital compliance teams. The number worth watching is not adoption but whether anyone publishes outcome data, because a PLOS Digital Health study this year found only 3 of 1,357 FDA-authorised AI medical devices had been evaluated on patient-centred outcomes. Convenience for clinicians is easy to demonstrate. Benefit to patients is not.
12. Broadcom AI Chip Revenue Jumps 221 Percent to $16.7 Billion
Broadcom reported $16.7 billion in AI semiconductor revenue for the third quarter of fiscal 2026, up 221 percent year over year, and guided to $21.7 billion for the fourth quarter. That would represent roughly 30 percent sequential growth on an already tripled base.
Broadcom's AI business is built on custom accelerators designed with individual hyperscalers rather than on general-purpose GPUs, which is why its growth rate exceeds Nvidia's even at a smaller base. Google's TPU line runs through Broadcom, and OpenAI's Jalapeno inference chip was manufactured with Broadcom on TSMC's N3P process. Guiding to $21.7 billion means Broadcom expects the custom silicon shift to accelerate rather than plateau, which is a specific bet that hyperscalers keep moving workloads off merchant GPUs.
My take: 221 percent growth in custom AI silicon is the clearest signal that the largest buyers are designing their way out of dependence on a single supplier. Set it beside Google taking a $12.2 billion warrant on Marvell tied to custom chip purchases and OpenAI building Jalapeno, and the pattern is unmistakable. Nvidia's $96.2 billion quarter is still enormous, and the second sourcing is now real rather than theoretical.
13. Microsoft Azure Crosses $100 Billion in Annual Sales
Microsoft disclosed fourth-quarter Azure revenue of $29.42 billion, up 42 percent year over year, with full-year Azure sales crossing $100 billion for the first time. The disclosure follows the company's earlier acknowledgement that it has approximately 2.2 million AI chips installed globally after spending around $280 billion since 2022.
A cloud business growing 42 percent at a $100 billion annual run rate is unusual, because growth rates normally decay as the base expands, and it reflects AI workloads arriving on top of existing cloud migration rather than replacing it. The tension in Microsoft's numbers remains the one Satya Nadella described earlier this year, where the constraint is powered buildings rather than chips. Revenue growth of 42 percent against a capacity constraint means the company is selling everything it can energise.
My take: the interesting question is what happens to that growth rate when capacity catches up, because right now demand exceeds supply and pricing reflects that. Broadcom's guidance and Dell's $95 billion backlog both point to supply arriving through 2027. If it arrives faster than demand grows, cloud AI pricing has further to fall than most models assume.
14. DOJ Backs OpenAI's Fair Use Defence Against the New York Times
The Department of Justice filed in support of OpenAI's fair-use defence in the New York Times copyright litigation on September 2, 2026. It is an unusual intervention, since the federal government does not routinely take positions in private copyright disputes between a publisher and a technology company.
The case is one of the most consequential open questions in AI, because a ruling against fair use for model training would apply retroactively to every model trained on scraped text, which is all of them. Government support does not decide the outcome and it does shape how a court weighs the public interest argument, which is one of the four statutory fair-use factors. It also lands the same week ByteDance settled its generative video dispute with the Motion Picture Association through a memorandum with no licensing fees, and those two paths, litigation and negotiated truce, are the only two on offer.
My take: I would not read the filing as a prediction of the result. What it does signal is that the administration views training data access as a competitiveness issue rather than purely a rights issue, which is a meaningful framing shift. For anyone building on these models, the practical exposure is unchanged for now, and a ruling either way would be the single largest legal event in the industry's history.
15. AI Model Prices in September 2026: Full Comparison
Here is where model pricing stands as of September 3, 2026, with cost per task included where Artificial Analysis publishes it.
The cost-per-task column is the one to read where it exists. Muse Spark 1.3 at $0.55 against GPT-5.6 Sol at $0.95 is a 73 percent premium on the OpenAI model for one index point of difference, and per-token pricing alone would not have told you that. Note also that three promotional rates expire between now and January: Claude Sonnet 5's ended August 31, GPT-5.6 Sol's ends in November, and Gemini Flash's ends December 31.
Choosing today: Muse Spark 1.3 if cost per completed task decides, Gemini 3.8 Flash if latency and price decide, Claude Opus 5 if general capability decides, Claude Fable 5.1 if your agentic prompts are cache-heavy, and Tencent Hy4 preview if you need frontier-scale open weights under a permissive licence. Detail sits in our best AI models ranking and the GPT-5.6 review.
16. What to Watch Next in AI
Four things carry into next week.
Whether Muse Spark 1.3's $0.55 cost per task holds up on independent workloads, and whether other labs start publishing cost per task rather than cost per token.
Fairwind Program access criteria, specifically whether volunteer open-source maintainers can realistically get through the vetting that Google says they qualify for.
The New York Times case timeline now that the DOJ has filed, and whether other agencies or amici follow.
Gemini 3.5 Pro, still unshipped, under Koray Kavukcuoglu's new leadership at Google DeepMind.
The through-line for September 3 is that the model market has stopped being a single race. Four labs shipped inside 48 hours and not one of them was competing for the same crown: OpenAI on autonomous capability, Anthropic on agentic price-performance, Google on security and speed, Meta on efficiency per task. If you are still choosing one model for everything, this was the week that stopped making sense. The right architecture now routes between four vendors, and the tooling to do that finally exists.
Frequently Asked Questions
What is Meta Muse Spark 1.3 and how much does it cost?
Muse Spark 1.3 is Meta's fourth Muse Spark release in five months, available through Muse Code and the Meta Model API with a 1 million token context. Pricing is unchanged from Spark 1.2 at $1.25 per million input tokens, $0.15 per million cached input, and $4.25 per million output. Meta reports roughly 20 percent fewer tool calls and 25 percent fewer tokens than Spark 1.2.
Is Muse Spark 1.3 better than GPT-5.6 Sol?
On cost efficiency, clearly. Muse Spark 1.3 (xhigh) costs $0.55 per Artificial Analysis Intelligence Index task against $0.95 for GPT-5.6 Sol (max), a premium of more than 70 percent on the OpenAI model. On raw capability they are close, with Spark 1.3 (max) scoring 62 on the Intelligence Index against Claude Opus 5 at 63.
What is Gemini 3.8 Flash Cyber?
Gemini 3.8 Flash Cyber is a security-focused variant of Google's Gemini 3.8 Flash, launched September 2, 2026 for vulnerability detection and automated patching. It scores 47.2 percent pass@1 on CWE-Bench against 47.8 percent for a leading frontier model at lower cost, exceeds 70 percent success across 20 programming languages internally, and produced 2.6 times more correct Chrome patches than comparable commercial models.
Which AI model leads CyberGym in 2026?
Gemini 3.8 Flash Cyber leads CyberGym at 86.2 percent as of September 2026, ahead of GPT-5.5-Cyber at 85.6 percent, Claude Mythos 5 at 83.8 percent, GPT-5.6 Sol at 83.6 percent, and Gemini 3.5 Flash Cyber at 77.5 percent. It displaced Z.ai's GLM-5.3, which had led at 84.5 percent and remains the only model in that group with published open weights.
What is Google's Fairwind Program?
Fairwind is Google's access programme for Gemini 3.8 Flash Cyber, restricting the security-focused model to government authorities, critical infrastructure operators, and software maintainers. It is not available through the general Gemini API, mirroring OpenAI's decision to gate Astra's advanced cyber capabilities behind partner vetting.
How much does Gemini 3.8 Flash cost?
Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output through December 31, 2026, the same introductory rate Gemini 3.7 Flash carried. On January 1, 2027 those rates double to $1.50 and $7.50, so 2027 budgets should be modelled at the higher figure.
Which is the best Chinese AI model in September 2026?
Kimi K3 from Moonshot AI leads at a composite score of 79.9, ahead of Alibaba's Qwen3.8-Max at 78.6 and Tencent's Hy4 preview at 78.3. The spread is narrow enough that weighting choices affect the order, and Hy4 preview achieves 78.3 at 770 billion parameters against Kimi K3's 2.8 trillion, under a more permissive Apache 2.0 licence.
Can ChatGPT access Epic medical records?
Yes, on a read-only basis. OpenAI launched an Epic integration for ChatGPT Health on September 1, 2026, letting authorised clinicians import appointment notes, laboratory results, medications, and specialist documentation covering more than 325 million patients. The model cannot write back to records. UCSF Health is a pilot partner.
How much AI revenue did Broadcom report?
Broadcom reported $16.7 billion in AI semiconductor revenue for the third quarter of fiscal 2026, up 221 percent year over year, and guided to $21.7 billion for the fourth quarter. Its AI business is built on custom accelerators designed with individual hyperscalers, including Google's TPU line and OpenAI's Jalapeno inference chip.
Did the DOJ back OpenAI against the New York Times?
Yes. The Department of Justice filed in support of OpenAI's fair-use defence in the New York Times copyright litigation on September 2, 2026. The intervention is unusual, since the federal government does not routinely take positions in private copyright disputes, and it does not determine the outcome.
Recommended Blogs
● OpenAI Astra Hits Critical Cyber Risk: AI News September 2 2026
● DeepSeek V4-Flash-Vision Goes Open Under MIT: AI News September 1 2026
● Tencent Opens a 770B Model Under Apache 2.0: AI News August 31 2026
● GLM-5.3 Weights Ship With a Hyperscaler Catch: AI News August 30 2026
● Best AI Models July 2026: Ranked by Use Case and Price
● GPT-5.6 Review: Sol, Terra, Luna Benchmarks and Pricing
● Kimi K3 Review: Benchmarks, Pricing, and K2 Comparison
Resources & Community
Join our community of 70,000+ AI enthusiasts and learn to build powerful AI applications! Whether you're a beginner or an experienced developer, Build Fast with AI helps you understand and implement AI in your projects.
● Website: buildfastwithai.com
● LinkedIn: Build Fast with AI
Agentic AI Launchpad 2026
A structured 6-week cohort program that takes you from AI basics to building and deploying real-world agentic AI systems. Includes live sessions, expert mentorship, project reviews, and a builder community network.
Ready to go from learning to building? Join the next cohort: Agentic AI Launchpad 2026
Free AI Resources
Access free tools, workshops, and micro-learning to keep building:
● AI Workshops: Free resources, upcoming events, and past recordings
● Unrot: Learn AI in 5 minutes a day (free micro-learning app)
● Gen AI Experiments: free cookbooks and notebooks on GitHub
Fairwind access criteria and independent Muse Spark 1.3 testing both land this week. Follow Build Fast with AI so each recap reaches you before your standup.
References
● Meta debuts Muse Spark 1.3 (Axios)
● Muse Spark 1.3 intelligence and price analysis (Artificial Analysis)
● Muse Spark 1.3 coding and agentic gains (Neowin)
● Gemini 3.8 Flash Cyber hunts vulnerabilities (VentureBeat)
● Gemini 3.8 Flash Cyber model page (Google DeepMind)
● Fairwind Program (Google DeepMind)
● Gemini 3.8 Flash rollout (9to5Google)
● ChatGPT Health adds Epic integration (TechCrunch)
● Best Chinese AI models September 2026 (BenchLM)
● Top tech news September 2 2026 (Tech Startups)
● Independent model evaluations (Artificial Analysis)
