- Services
The Frontier Price Collapse
By Tobias Wupperfeld · Published Frontier providers opened a direct price war in September. Anthropic released Claude Opus 5.5 at 4 dollars per million input tokens. OpenAI answered with GPT-6 Sol at 2 dollars and GPT-6 Luna at 10 cents.
The model count in our database rose to 862 as monthly research submissions surged 24 percent. Google announced Gemini 4 Argon for enterprise workflows. SpaceXAI released Grok 4.7 with longer execution capacity at flat pricing.
Autonomous tooling expanded across the market. OpenAI introduced its managed Agents API. MongoDB launched Atlas Agent Engine for persistent memory. Nvidia agreed to acquire Hugging Face for 12.9 billion dollars.
September was an intense month in AI development. Anthropic opened with Claude Fable 5.1 and a gated Mythos tier. They followed with Claude Opus 5.5 and Sonnet 5.5. I guess they decided Mythos is not going to end the world. Cutting Opus 5.5 pricing to 4 dollars per million input tokens shifts enterprise budgeting immediately.
OpenAI responded rapidly. They launched GPT-6 Astra, followed by Sol at 2 dollars and Luna at 10 cents per million input tokens. Frontier performance is no longer priced as a scarce commodity. Major vendors are trimming margins to win default placement in company software stacks. For enterprise buyers, the ROI on complex planning pilots just improved dramatically.
Our client engagements at Made By Agents are already reflecting these shifts. I do not think coding models need plan mode anymore. We are thinking of killing plan mode and using the shift plus tab hotkey to adjust effort levels. Add-ons like rtk and ponytail are proving practical in daily work. In addition, Opus 5.5 can generate accurate SVG logos and graphics programmatically.
The scientific announcements this month were striking. Anthropic reported that Claude agent swarms identified a CRISPR-like enzyme system in bacteriophage DNA. OpenAI announced that an internal model produced a verified proof for the Navier-Stokes problem in Lean. These are not basic chat demos. Autonomous systems are starting to resolve open scientific problems without human direction.
Politicians in Washington and London are responding to this speed. Lawmakers proposed ban legislation in both the United States and the United Kingdom. Technology executives entered an open public debate over development pacing. Meanwhile, Nvidia agreed to buy Hugging Face for 12.9 billion dollars. Technology leaders should bypass political noise and focus on real unit economics.
We rebuilt the Made By Agents website this month to help teams monitor these movements. We launched our capex tracker and our hiring index. We also expanded our pricing indices across more GPU rental marketplaces. These tools provide hard numbers for your roadmap before you commit to vendor contracts.
The launches, models, and research most worth a builder's attention this month, with a link to the source.
Google announced Gemini 4 Argon on September 30, 2026, targeting complex workflows across software engineering, legal, finance, and cybersecurity. The model rolls out first to selected security teams through the Fairwind Program. I view this as Google prioritizing high-value enterprise processes over consumer chat interfaces.
Source: Google
SpaceXAI released Grok 4.7 on September 21, 2026, keeping the exact price and speed of Grok 4.6. The larger base model was trained with longer reinforcement learning to tackle multi-hour problems and benchmarked at the frontier on CursorBench 4.0. Vendor costs stay flat while your automated workflows gain the ability to sustain longer execution times.
Source: SpaceXAI
The Institute of Foundation Models released K2 Horizon on September 3, 2026, spanning six open models from 0.9 billion to 375 billion parameters. The release provides open weights, training data recipes, code, and intermediate checkpoints across the entire training lifecycle. This gives companies an auditable foundation stack that reduces vendor lock-in risk for private deployments.
Source: Institute of Foundation Models
OpenAI launched GPT-6 Astra on September 4, 2026, positioning the system as its flagship model for coding, science, mathematics, and computer interaction. I recommend running initial pilots on Astra for complex planning tasks before routing routine production calls to cheaper tiers.
Source: AI Front Page
OpenAI launched its Agents API in public beta on September 10, 2026, giving developers access to the cloud harness and infrastructure that powers Codex. The system manages context, files, code execution, and subagent coordination for tasks that run across multiple days. Teams can retire custom agent infrastructure and reallocate roadmap budget toward product workflows.
Source: OpenAI
MongoDB launched Atlas Agent Engine on September 29, 2026, adding execution, persistent memory, and governance layers for enterprise AI systems. This solves the production memory bottleneck and cuts architectural complexity by keeping agent state inside your primary database.
Source: MongoDB
NVIDIA launched the Open Agent Safety Platform on September 28, 2026, combining OpenShell software with Sentry reference designs. The platform establishes security guardrails and control across agent runtimes, compute clusters, and robotics systems. This reduces operational risk for organizations preparing to deploy autonomous software across sensitive networks.
Source: NVIDIA
Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 output on September 22, 2026, while OpenAI launched GPT-6 Sol at $2 input and $10 output. OpenAI also introduced GPT-6 Luna at $0.10 input and $0.50 output per million tokens. I guess Anthropic decided Mythos is not going to end the world, and this price war immediately lowers production costs for agent builders.
Source: Roundup
OpenAI announced on September 8, 2026, that an internal model produced a proof resolving the 90-year-old Navier-Stokes existence and smoothness problem. The company published both the mathematical writeup and a formal verification in Lean. This breakthrough proves autonomous AI systems can generate verifiable discoveries in fundamental science without human guidance.
Source: OpenAI
Meta launched Meta Enterprise Platform on September 28, 2026, establishing a major enterprise pillar to deliver models, agents, and infrastructure. This provides enterprise technology leaders with a direct commercial vendor option to compete against Microsoft and Google cloud contracts.
Source: Meta
The total number of tracked models grew from 753 to 862 in September. Frontier reasoning benchmarks saw fresh gains. Claude Opus 5.5 pushed the top score on Humanity's Last Exam to 61.4, up from 55.5 in August. Scores on standardized tests like GPQA and SWE-Verified remained steady at 94.1 and 80.6.
Enterprise buyers should focus on these specialized reasoning gains rather than general chat benchmarks. High-difficulty evaluations reveal which systems can handle complex multi-step analysis without failing. Run initial pilots on high-scoring models before deciding where to route production traffic.
Tracking 862 hosted models across OpenRouter, ppq.ai, and kie.ai.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| OpenAI GPT-5 | $0.02 | $0.20 |
| OpenAI o-series | $0.55 | $2.20 |
| Anthropic Claude Opus | $1.43 | $7.15 |
| Anthropic Claude Sonnet | $0.80 | $4.00 |
| Google Gemini 2.5 Pro | $0.38 | $3.00 |
| Google Gemini Flash | $0.05 | $0.20 |
Median input prices fell from 0.50 dollars to 0.454 dollars per million tokens. Median output prices held steady at 1.929 dollars. Several utility models saw aggressive price drops. Z.ai GLM Latest fell by 94.87 percent, while Ling 3.0 Flash dropped 93 percent.
At the frontier, OpenAI released GPT-6 Sol at 2 dollars input and GPT-6 Luna at 0.10 dollars input. These aggressive cuts reduce the operational risk of multi-agent loops. Technology leaders should audit their vendor bills and route background tasks to these discounted tiers.
Gemma 3 27B IT: cumulative cost over 12 months
The API saves about $264 over a year at this workload.
Local vs API assumes 50M tokens a month, split evenly between input and output, with the cheapest cloud GPU that fits the model billed at 8 hours a day.
Renting GPUs continues to carry higher monthly costs than cloud APIs for low-volume workloads. Running Gemma 3 4B IT via API costs 4 dollars per month based on our benchmark volume. Renting an RTX A4000 on Vast.ai to host the same model costs 20 dollars per month. Gemma 3 27B IT costs 13 dollars via API versus 35 dollars on a rented RTX 3090.
Self-hosting only delivers positive ROI when hardware runs near full utilization 24 hours a day. Unless strict privacy rules require private infrastructure, managed APIs remain the lower cost option for initial pilots.
The most-starred open agent frameworks, with how their stars and downloads moved.
Developer activity showed clear consolidation around production-ready orchestration. LangGraph led the market with 10,250,952 weekly downloads and gained 11 percent in GitHub stars to reach 39,709. LangChain maintained steady traffic with 3,401,959 weekly downloads and 140,318 stars.
Alternative frameworks saw flat adoption. AutoGen held at 86,148 downloads, while CrewAI recorded 569,590. Teams building agent workflows are choosing graph-based control over open-ended multi-agent chat loops. We advise clients to standardize on deterministic state graphs for business logic.
New papers posted to arXiv in the categories that drive applied AI. A rough gauge of how fast the field is moving, not a quality measure.
13,150 papers in total, up 24.1% on the month. Papers land in more than one category, so the four numbers above add up to more than the real total.
Research output surged sharply in September. Monthly paper submissions grew 24.07 percent to reach 13,150 papers on arXiv. Machine learning expanded the fastest, rising 39.84 percent to 5,437 publications. Artificial intelligence grew 19.50 percent to 5,784 papers.
This surge reflects intense competition in autonomous agent design and formal mathematical reasoning. The volume of published work means internal teams cannot track every new paper. Focus roadmap investments on reproducible implementations rather than academic preprints.
A point-in-time read of major provider uptime, taken when this report was generated.
Reliability data highlighted real operational risk among leading providers. OpenAI reported 24 incidents during the month, leaving its overall status degraded. Anthropic recorded 13 incidents while maintaining an operational rollup status. Cursor logged 31 incidents. Google AI Studio stood out with zero recorded incidents.
Production systems cannot rely on a single model endpoint. Downtime in your provider harness halts client workflows and damages customer trust. Every production deployment should include automated fallback routing across multiple vendors.
I am watching the enterprise rollout of Google Gemini 4 Argon next month. Google is targeting legal, financial, and cybersecurity teams through its Fairwind Program. If Google captures mission-critical enterprise workflows, Anthropic and OpenAI will need to adjust their enterprise roadmaps.
I am also monitoring how developers receive new tooling from OpenAI DevDay and Meta. OpenAI previewed Dots, while Meta introduced Muse. At the same time, SpaceXAI released Grok 4.7 to coordinate multi-agent teams inside dedicated virtual machines.
Finally, we will watch the closing steps of the Hugging Face acquisition by Nvidia. A 12.9 billion dollar deal places the primary open model hub under the dominant chip maker. We will see whether open-source communities migrate workloads to independent registries.
Input tokens are the words and code you send into a model. Output tokens are the words the model generates in response. Providers charge significantly more for output tokens because generating text requires more compute power. Budget forecasts should account for this price difference when planning agent loops.
Running a local model only pays off when your workload saturates a GPU around the clock. Light or variable usage makes cloud APIs far cheaper because you only pay for completed requests. Hosting your own hardware also introduces maintenance overhead, cooling costs, and software stack management.
Humanity's Last Exam measures model accuracy on advanced, closed-ended questions across diverse scientific disciplines. It tests graduate-level reasoning in mathematics, physics, and humanities subjects. A high score shows a model can execute complex analytical tasks without relying on memorized answers.
Chat models produce conversational sentences that software systems must parse and clean. Decision models return concise structured outputs such as a numerical score or a single choice. This design cuts latency, eliminates parsing bugs, and lowers token costs in automated business workflows.
A managed agent API provides hosted cloud infrastructure to coordinate autonomous AI agents. The vendor manages server execution, file storage, memory state, and task handoffs across long sessions. This setup removes the need to maintain custom server harnesses but links your architecture directly to a single provider.
One email the day each edition is out, plus The Agent Roundup, our weekly note for AI builders. Free, unsubscribe any time.
