OpenAI Buys Anthropic's Customers With 90% Discounts as Microsoft and AWS Open-Source the Agent Stack
OpenAI Buys Anthropic's Customers With 90% Discounts as Microsoft and AWS Open-Source the Agent Stack
Sunday, 25 October 2026 | Coverage window Since 24 October, 18:44 IST | Compiled 25 October, 18:45 IST
The agent stack commoditized on two fronts at once: OpenAI is buying coding workloads from Anthropic with discounts of up to 90 percent, while Microsoft and AWS open-sourced the harness and runtime layers for agent swarms, answering Google's agent infrastructure from earlier in the week. The open ecosystem kept pace - vLLM matched DeepSeek's own engine, Cerebras showed the arithmetic of serving 3 trillion parameters across 59 wafers, and Nvidia open-sourced the data engine feeding its robot models - while an 11-hour unattended run let Gemini 4 Argon improve Google's own agent framework, with humans keeping the merge keys.
1. OpenAI's discount offensive pulls Replit, Rippling and DataCamp workloads from Anthropic
REPORTING
The Information reported on 24 October that OpenAI is winning Anthropic's largest customers with steep discounts. Replit signed a contract worth more than $100 million a year after OpenAI offered over 30% off. Rippling CEO Parker Conrad said OpenAI's share of its coding workload rose from 16% to 52% while Anthropic's fell from 57% to 18%. DataCamp's CEO said OpenAI pricing cut costs 80-90%, and a US bank with over $7 billion in assets had its ChatGPT seats doubled to 500 employees at under $7 per user. Anthropic's growth slowed from Q2 to Q3, per an investor memo cited in the report.
ANALYSIS
The frontier price war is now quantified, and it is being fought on Anthropic's best turf: coding. OpenAI is partly buying share at the expense of margin, but real efficiency gains in its Sol and Luna lines fund part of the cuts. Replit's switch to OpenAI as default was covered 17 October; this is the dollar account of what that decision cost Anthropic. Both labs also face Chinese open models delivering comparable coding performance for a fraction of the price, which compresses everyone's floor.
The Information / The Decoder · 24-25 Oct 2026 · The Information · The Decoder report
2. Microsoft open-sources Agensh, a leaderless harness that scales to 1,024 agents
REPORTING
Microsoft Research open-sourced Agensh on 24 October: a decentralized multi-agent harness with no central orchestrator, where workers claim sub-tasks, share findings over a message interface and merge code into a shared Git repository. The release includes the harness, launch scripts and a reproduction guide for the 22 September preprint. On the five hardest ProgramBench tasks with GPT-5.6-sol, a six-hour budget and no internet, mean test-pass rate rose from 19.31% with one agent to 28.78% with 128 agents, roughly a 49% relative gain; on Pandoc it climbed from 33.89% to 50.94% at 128 agents and 55.06% at 1,024.
ANALYSIS
The claim that matters is organizational, not numerical: peer coordination, integration workflows and specialized roles emerged without being programmed. Microsoft is equally honest about the ceiling - going from 128 to 1,024 agents bought little, which tells the industry where swarm economics stop paying. With Google's Agent Compute Engine (22 October) and AWS's Infinity (story 3), the big three clouds have now all published their reference designs for agent swarms in the same week.
Microsoft Research / MarkTechPost · 24-25 Oct 2026 · GitHub repository · Microsoft Research publication
3. AWS open-sources Infinity, a runtime that fits 75,000 waiting agents in 8 GB
REPORTING
AWS released Infinity under Apache 2.0 on 23 October, built with UC Berkeley in the Hydro project. Infinity treats an agent like a stackless coroutine: when the model calls a tool, the runtime persists the conversation, suspends the agent and frees the process, resuming it from saved state when the result arrives. In a benchmark with a simulated model, 80,000 agents ran twenty turns each - 3.2 million completions - at roughly 100 KB per idle agent, so about 75,000 concurrent agents fit in the 8 GB of a Raspberry Pi 5. A serverless reference deployment runs on SQS FIFO, Lambda, Aurora DSQL and DynamoDB, and synthetic tool calls deliver asynchronous events without breaking the one-result-per-call rule. AWS calls it an early research release.
ANALYSIS
Read with Google's Agent Compute Engine two days earlier, the industry has decided agents are a workload class of their own, and the expensive part is the waiting, not the thinking. Google isolates the machine each agent runs on; AWS eliminates the process while it waits. Both designs assume fleets of long-lived agents, which makes this week's monitoring and containment stories the other half of the same bet.
AWS Open Source Blog · 23 Oct 2026 · AWS blog post · GitHub repository
4. Cerebras shows how it spreads a 3-trillion-parameter model across 59 wafers
REPORTING
Cerebras published a detailed account on 22 October of deploying trillion-parameter frontier models on its WSE-3T wafers, each with 44 GB of on-chip SRAM, using the open-source Kimi K3 (about 3 trillion parameters, 69 linear-attention and 24 dense layers) as the worked example. With 4-bit weights and an 8-bit KV cache, each dense layer gets one wafer for its KV cache while two linear-attention layers share one, mapping K3 across 59 wafers and serving it at 1,000-2,000+ tokens per second. Cerebras notes K3 scores 88.3% on Terminal-Bench 2.1 against GPT-5.6 Sol's 88.8%.
ANALYSIS
Wafer-scale serving is now legible engineering rather than marketing: the arithmetic of layers, SRAM and KV cache is public, and linear attention's fixed-size state is what makes on-chip residency of a 3T model feasible at all. Frontier open weights being served at thousands of tokens per second on non-Nvidia silicon is also the supply-side answer to the price war in story 1 - cheap tokens need somewhere to come from.
Cerebras · 22 Oct 2026 · Cerebras blog post · Technical walkthrough
5. Nvidia open-sources Flora, the data engine labeling robot video for its GR00T models
REPORTING
Nvidia's GEAR lab open-sourced Flora, a suite of three fine-tuned Qwen3.5 models - a 9B segmenter, a 27B captioner and a 27B rewarder - that label robot video with timestamped captions, step-by-step progress ratings and success judgments. Nvidia estimates labeling one hour of video costs $0.72 with Flora against $88.49 with GPT-6 Astra, at slightly higher quality (57.1% versus 53.7%), and says Flora already annotates up to 200,000 hours of video per day on 512 GB300 GPUs for GR00T training. The Decoder notes the costs are projections at an assumed $10 per GPU-hour, that Astra still leads on WGO-Bench and robot-video captioning, and that Flora falsely flags about 11% of normal steps as errors. Models are on Hugging Face.
ANALYSIS
The capability that matters is catching mid-episode failures: Flora-Rewarder detects 53.3% of failed steps where no competitor exceeds 11.7%. Annotation was robotics' human bottleneck, and Nvidia just open-sourced its own answer - which is also an ecosystem play, since every lab that adopts Flora's pipeline lands on Nvidia silicon. The honest limitations, published alongside, are a model for how vendor claims should ship.
Nvidia GEAR / The Decoder · 25 Oct 2026 · Nvidia research page · The Decoder report
6. vLLM matches DeepSeek's DeepInfer and scales DeepSeek-V4.1-Flash across a full GB300 rack
REPORTING
The vLLM team reported on 23 October that with prefill-decode disaggregation and Wide-EP it serves DeepSeek-V4.1-Flash on one NVIDIA GB300 NVL72 rack at 954K tokens per second per GPU on SemiAnalysis AgentX, at a P90 of 196 tokens/s per user and $38.96 per GPU-hour at DeepSeek's off-peak prices - 7.9 times its single-node result from 2 October. Decode throughput lands within 6% of DeepInfer, DeepSeek's own engine released this week, across batch sizes 8-256, and a specialized decode path reaches 826 tokens/s for a single user. The work builds directly on kernels DeepSeek open-sourced, including MegaAttn, MegaMoE and Mega-mHC.
ANALYSIS
DeepInfer's 21 October launch argued only the model's author could serve it efficiently; two days later a general-purpose engine matched it, with online serving features DeepInfer lacks. The lineage runs both directions - DeepInfer was originally built on vLLM - which is the open ecosystem working as intended. For operators, the takeaway is that frontier-class open serving no longer requires adopting a vendor-specific stack.
vLLM Blog · 23 Oct 2026 · vLLM blog post · DeepInfer vs SGLang comparison
7. An 11-hour unattended Gemini 4 Argon run improved Google's own agent framework
REPORTING
Google engineers let Gemini 4 Argon run unattended for 11 hours on the Agent Development Kit - over 200,000 lines of Python across about 800 files with nearly 17,000 unit tests - using Antigravity's /goal mode and repository-specific skill files as an onboarding guide. InfoQ covered the writeup on 24 October. The agent established performance baselines, proposed changes, reviewed its own work and ran all 16,892 unit tests; two optimizations were accepted, cutting in-memory session-preparation overhead by about 40% in long sessions and cold package-import time by 21%, trimming a 158-turn internal run from 65.6 to 56.1 seconds. Four other proposed changes were rejected on developer review for reliability and maintainability concerns.
ANALYSIS
This is the self-improvement story with its governance intact: the model proposed, humans disposed, and only a third of the changes survived review. The reusable artifact is the scaffolding - skill files plus a persistent goal prompt - which any team can copy, not the two patches. That a frontier model did cross-module legwork a profiler cannot do, tracing callers across 40 files, is the quiet capability signal.
Google Developers Blog / InfoQ · 24 Oct 2026 · InfoQ report · Google Developers Blog
8. Reservoir seeks roughly $340 million from Suno and Anthropic over training on its catalog
REPORTING
Music publisher Reservoir filed two suits on 21 October in the Northern District of California, reported by Music Business Worldwide on 22 October. Against Suno it lists 1,050 compositions - including Joni Mitchell's Both Sides Now, Johnny Cash's Ring of Fire and Madonna's Papa Don't Preach - with pleaded figures just over $186 million including DMCA counts; a separate suit against Anthropic covers 853 compositions, with pleaded figures around $151 million. The complaints allege scraping via Bright Data proxies from YouTube Music, Deezer and Genius, citing leaked Suno code reported by 404 Media, and seek injunctions plus destruction of unauthorized copies including those in training datasets and model weights. Reservoir is represented by Richard Busch, who won the Blurred Lines case.
ANALYSIS
The escalation is the destruction demand reaching into model weights, and the use of Suno's own licensing deals with Warner, BMG and Believe as evidence that a training-data market existed to be licensed. Naming Anthropic alongside Suno signals publishers now see model labs, not just music generators, as infringers. With Bartz v. Anthropic and Thomson Reuters v. ROSS as cited precedent, the fair-use ground is narrowing for anyone who trained on scraped catalogs.
Music Business Worldwide / Justia · 21-22 Oct 2026 · MBW report · Suno docket
9. Pennsylvania becomes one of the first US states to require labels on AI-generated ads
REPORTING
Pennsylvania Governor Josh Shapiro signed the AI Deceptive Advertising Act on 22 October, CBS and The Verge reported. The law requires a clear disclosure on AI-generated or heavily edited advertising content that could mislead buyers about a product, takes effect in 60 days - 21 December - and was strengthened by a June amendment that removed its knowledge requirement, so advertisers cannot dodge liability by claiming ignorance. The Verge notes the act leaves significant loopholes.
ANALYSIS
Absent federal rules, US AI disclosure law is being written state by state, and Pennsylvania's version puts the compliance burden on ad platforms and buyers rather than model labs. A December effective date means provenance and labeling plumbing becomes an ad-tech requirement this quarter, and other statehouses now have a template with a governor's signature on it.
The Verge / CBS · 22-24 Oct 2026 · The Verge report · CBS Pittsburgh report
10. KRAFTON releases Raon-1 mini, a permissively licensed 30B MoE with a matched draft model
REPORTING
KRAFTON AI released Raon-1 mini preview on Hugging Face under Apache 2.0 (access granted on request): a Mixture-of-Experts model trained from scratch with about 28.8 billion total and 4.3 billion active parameters, a 262,144-token context and strength in Korean conversation and agentic tasks. It ships with a speculative-decoding draft model that KRAFTON says reaches up to 2.18 times higher output throughput on a single NVIDIA B300 - a vendor-reported figure. Support for Transformers CPU offload and RTX 5090 inference landed within days.
ANALYSIS
A games company funding a from-scratch open model family is a newer pattern, and the packaging is the point: a long-context MoE plus its own draft model gives self-hosters a single-GPU drop-in without a cluster. In a week when OpenAI is buying workloads with discounts, permissively licensed models that run on one card are the zero-marginal-cost alternative at the bottom of the market.
KRAFTON / TPS · 22-25 Oct 2026 · Hugging Face model card · Dosa brief
Ten distinct developments, none covered in prior editions. Four primaries predate the window but were missed at the time and appear as first coverage: AWS Infinity (23 October), the Cerebras deployment post (22 October), the Reservoir complaints (filed 21 October, reported 22 October) and Raon-1 mini (22 October). Replit's switch to OpenAI as default was covered 17 October; story 1 is The Information's quantified account of the price war behind it, a distinct development. Flora's cost figures and Raon-1's throughput gains are vendor projections, labeled in the stories; Agensh and Infinity results are measured against simulated models or single benchmarks. No public X post met the evidence threshold this cycle, so no Firsthand signal item is included. Weighed and dropped: Jack Clark's self-improving-AI forecast shift (reported 14-15 October, stale), the ProjectDiscovery Qwen backdoor demonstration (7-16 October, stale), a dev.to digest framing Anthropic's October 2025 one-million-TPU Google deal as this week's news, the ARC Prize GPT-6 Astra harness rerun (September story recycled), Meta's Sundial acquisition (confirmed 23 October, minor), OpenMobile-2, Cognition's Devin Vulnerability Fixes plugin, and H200 rental-price data. Prior editions remain unchanged.
Coverage window Since 24 October, 18:44 IST | Compiled 25 October, 18:45 IST