TL;DR
- OpenAI cut API prices for GPT-6 Sol and Luna by 50% versus GPT-5.6 promotional rates. Sol: $2/$10 per million input/output tokens. Luna: $0.10/$0.50. On AutomationBench — Zapier's benchmark of 47-tool business workflows — GPT-6 Sol at highest effort scores 33.2% at $0.27 per task. Claude Opus 5 at max effort scores 26.9% at $3.00 per task. That is 11.1x the cost for lower benchmark performance. On DeepSWE v1.1, Sol at max effort scores 68.8% — within 1.1 points of Claude Fable 5's best of 69.9% — at 80% lower cost per task. - DeepSeek V4.1-Flash-1017 released this week hits 77% on DeepSWE v1.1 — beating Claude Opus 5, GPT-5.6 Sol, and coming within striking range of the top coding benchmarks — at under $0.20 per task. Search and Fetch APIs now open to all developers. - Nvidia released Nemotron 3.5 Super VL: 120B total parameters, 12B active (MoE), 1M-token context, multimodal (text/image/video), free on Hugging Face under OpenMDW-1.1. Architecture: Mamba-2-Transformer hybrid with C-RADIO vision encoder. 5x higher throughput for agentic tasks than competing open models in class. Scale AI, Ramp, and AI21 have post-trained it — reporting 40x lower inference cost, 10x lower latency, 40% higher accuracy on their specific workflows. Bolt 74B-A9B (24GB, fits on RTX Spark, 84 TPS) and Bolt E4B (4GB for Jetson/edge) extend the stack down to embedded deployments. - One week ago, Gartner projected $3.5 trillion in AI infrastructure at write-down risk by 2030, forecasting AI revenue of $3.6 trillion against infrastructure spending requiring $8.5 trillion to hit a 12% ROIC. The price cuts this week are the mechanism Gartner was describing. Revenue per API call is falling faster than infrastructure was underwritten to support.
Last week Gartner published the write-down math. This week the prices moved.
GPT-6 Sol at $0.27 per task. DeepSeek Flash at $0.20 per task. Nvidia giving away a 120-billion-parameter multimodal model for free. These are not promotional teaser rates — they are the new market prices for frontier-class AI capability in October 2026.
For enterprise teams, this is a planning moment. Budget models built on 2024-2025 API pricing — $15 per million input tokens for GPT-4, $30 for Claude 3 Opus — are wrong by two orders of magnitude. The question is not whether frontier AI is affordable. It is which of the providers currently charging 10x above market will survive the consolidation Gartner is forecasting.
The GPT-6 Sol and Luna pricing
GPT-6 Sol ($2/$10 per million input/output tokens, 90% cache discount on repeated input) and Luna ($0.10/$0.50) were cut 50% from GPT-5.6 promotional pricing. OpenAI attributed the reduction to improvements in caching and inference.
The benchmark numbers that matter for enterprise automation planning:
AutomationBench 1.0.6 (Zapier's 47-tool simulation of real SaaS business workflows, the most enterprise-relevant agentic benchmark currently public): - GPT-6 Sol at xhigh effort: 33.2%, $0.27/task - Claude Fable 5.1 + Opus 5 fallback at max effort: 31.4%, $2.40+/task (figure understates real cost — Opus 5 fallback fired on ~40% of tasks, not included in reported cost) - GPT-6 Astra at low effort: 30.3%, $1.05/task - Claude Opus 5 at max effort: 26.9%, $3.00/task
Sol scores higher than every Anthropic model on this benchmark at 9% of Opus 5's cost per task. The efficiency gap is not marginal.
DeepSWE v1.1 (software engineering): - GPT-6 Sol at max effort: 68.8% - Claude Fable 5: 69.9% at 80% higher cost per task - GPT-6 Luna at max effort: 66.6%, 93% cheaper than Opus 5, 96% cheaper than Fable 5
Agents' Last Exam (complex professional workflows): - GPT-6 Sol at max effort: 56.4%, above Opus 5's highest score, 60% lower cost per task - GPT-6 Luna at max effort: 66.6%, comparable to Opus 5 and Fable 5 at medium effort, 93% and 96% cheaper
OSWorld 2.0 computer use: - GPT-6 Sol at xhigh effort: 60.5%, approximately equal to Claude Opus 5 at medium effort (60.3%), 80% lower cost
One number OpenAI disclosed internally: daily token use at the median OpenAI researcher crossed $600; at the 90th percentile, $7,000. OpenAI is cutting prices while its own internal usage costs are at that scale. The inference economics have changed materially enough that a 50% cut remains profitable at these volumes.
GPT-6 Astra remains the strongest model for computer use per OpenAI's own positioning. At low effort, it scores 30.3% on AutomationBench — below Sol's 33.2% — at 3.9x Sol's cost per task. For most enterprise automation use cases, Sol is not a downgrade from Astra. It is a different optimization point on the cost-performance curve.
DeepSeek V4.1-Flash-1017
The incremental update is punchy enough to note. DeepSeek V4.1-Flash-1017 scores 77% on DeepSWE v1.1 at under $0.20 per task. That puts it above Claude Opus 5, above GPT-5.6 Sol, and within the top tier of coding benchmarks — at a price that undercuts even Luna on coding workloads. Search and Fetch APIs are now open to all developers, removing the waitlist restriction that had limited enterprise adoption.
DeepSeek's pricing model — inference at cost, funded by the lab's quantitative trading revenues — does not require the AI infrastructure returns that Gartner is modeling for neocloud viability. It is competing on a different economic basis than any US lab. This matters for enterprise procurement: DeepSeek pricing is not subject to the same revenue pressure that is driving the OpenAI and Anthropic price cuts.
Nvidia Nemotron 3.5 Super VL
The 120B-A12B (12B active per token) MoE model is free on Hugging Face. Architecture: Mamba-2-Transformer hybrid with MoE layers, C-RADIO vision encoder for text/image/video input, native Multi-Token Prediction for speculative decoding, NVFP4 quantization. 1M token context. Configurable reasoning mode.
What it is useful for: specialized agents running continuously on enterprise workflows. The design targets always-on agentic tasks — instruction following, tool use, multistep business workflows, visual document understanding. On Artificial Analysis Intelligence Index agentic tasks, it delivers over 600 tokens per second on Baseten endpoints. 5x higher throughput than competing open models in its size class.
Three organizations that have already post-trained it: Scale AI (specialized data-labeling agents), Ramp (finance automation), AI21 (document processing). Their reported outcomes: 40x lower inference cost, 10x lower latency, 40% higher accuracy. These figures are on their specific workflows after fine-tuning, not on general benchmarks — but they indicate the model's responsiveness to domain specialization.
The edge stack: Bolt 74B-A9B (24GB, 2-bit quantization via Star Elastic compression from the 120B parent, 84 TPS on RTX Spark, #1 open model on PinchBench in its class) for local agentic workloads. Bolt E4B (4GB for Jetson, GeForce RTX, DGX Spark) for conversational agents and physical AI. Together: data center to desktop to Jetson edge in one model family, with the same weights and architecture across the full deployment range.
Licensing: OpenMDW-1.1, not Apache 2.0. The digest cited Apache 2.0 — that's Reflection Beam 501B. Nemotron 3.5 Super VL uses OpenMDW-1.1, which permits commercial use but carries specific terms that enterprise legal teams will want to review before deploying in production.
What this week's pricing means for enterprise AI budgets
Three numbers for finance and procurement teams planning 2027 AI spend:
$0.27 per task (GPT-6 Sol, AutomationBench) vs $3.00 per task (Claude Opus 5, same benchmark) — that gap is 11x. If your 2026 automation budget was built on Opus 5 or equivalent pricing, Sol or Luna pricing cuts your cost by 80-93% for similar or better benchmark performance on business workflow tasks.
$0.20 per task (DeepSeek V4.1-Flash-1017, DeepSWE v1.1 coding) — for coding and software engineering workloads, DeepSeek Flash is now the price floor. Any provider charging more needs to justify it on capability, compliance, or data sovereignty grounds.
Free (Nemotron 3.5 Super VL) — the cost floor for on-premises multimodal agentic inference is zero for the weights. Deployment infrastructure (GPU cluster, serving layer, fine-tuning compute) is not free. But the model itself no longer has a license fee. For organizations with existing GPU infrastructure or sovereignty requirements that preclude cloud inference, the open-weight frontier has reached frontier-class multimodal capability.
The Gartner write-down math: infrastructure providers need $8.5 trillion in revenue through 2030 to cover costs. Each successive price cut reduces the revenue available to repay that infrastructure. OpenAI's 50% cut, DeepSeek's $0.20/task pricing, and Nvidia's free open-weight strategy are not isolated decisions — they are competing claims on an AI revenue pool that Gartner already projected as insufficient to support the infrastructure that serves it. The consolidation Gartner predicted (80% of neoclouds ceasing independent operations by 2030) accelerates each time the market price for inference drops.
For enterprise buyers, this is a favorable market dynamic in the short term: the cost of frontier AI capability is falling faster than almost any prior enterprise technology category. The risk is counterparty stability — the neocloud GPU rental agreements and multi-year API commitments signed at 2024-2025 pricing are being underwritten by suppliers whose revenue models are under the same pressure the price cuts create. Anthropic's $518 billion in non-cancellable compute commitments (disclosed in the S-1) and OpenAI's own infrastructure obligations are both being serviced while API prices are cut in half.
The enterprises that win in this environment: those that standardize on portable abstractions (LiteLLM, before the supply-chain attack demonstrated its risks; or equivalent proxy layers with vetted, pinned dependencies), maintain the ability to switch between providers on pricing signals, and have not locked into long-term agreements at a single provider's 2024 prices. The enterprises that lose: those that signed 3-5 year compute contracts with neoclouds at peak 2025 rates, or made architectural bets on a single proprietary model family that is now being outperformed at 9% of its cost.
