Notion Rebuilt Its AI Agent to Run for 24 Hours Instead of 5 Minutes. Every Activation Metric Your Product Team Tracks Is Now Wrong.

On September 29, 2026, OpenAI stood on a DevDay stage and announced a model priced at one-fifth the cost of its own best model. GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Astra — the model just above it in OpenAI's tier structure — costs $10 input and $50 output. The price difference is not a temporary promotion or an introductory offer. It is a permanent tier structure that OpenAI is calling the new normal.

The headline claim from OpenAI was "near-Astra intelligence for a fifth of the price." That claim deserves scrutiny — the benchmarks are self-reported, the use-case gap between Sol and Astra is real for some workloads, and enterprise buyers should test rather than trust marketing language. But the pricing structure itself is unambiguous: OpenAI has publicly defined a near-equivalent option at a permanent five-to-one discount from its flagship tier. That is not a discounting strategy. It is a commoditization signal.

Every enterprise that approved an AI budget in Q3 2026 based on Astra-tier pricing assumptions now has a recalculation to do. Every AI vendor whose pricing is anchored to frontier model costs has a margin problem. And every product team that has been waiting for AI to become cheap enough for high-volume workloads just got the answer they were waiting for.

The Pricing History: From GPT-4 to GPT-6.1 Sol

To understand what GPT-6.1 Sol represents, it helps to trace the cost curve that got here.

GPT-4 Turbo, released in November 2023, was priced at $10 per million input tokens and $30 per million output tokens — figures that made AI-powered products economically viable for carefully scoped use cases but prohibitive for high-volume applications. A company running a million customer interactions per month, each requiring 2,000 input tokens and 500 output tokens, was looking at roughly $20,000 per month in inference costs at GPT-4 Turbo pricing. That math pushed most enterprise AI implementations into narrow, high-value use cases where the cost was obviously justified.

The cost curve began moving seriously in 2024, as inference optimization, hardware efficiency improvements, and competition from Anthropic and Google drove prices down. By mid-2025, comparable capability to GPT-4 Turbo was available at roughly 10-20% of the original price. By early 2026, the trend had continued: the launch of DeepSeek's cost-compressed approach demonstrated that near-frontier performance was achievable at a fraction of previous compute requirements, forcing every major provider to reprice.

GPT-6 Sol, released earlier in 2026, already represented a meaningful price reduction from Astra. GPT-6.1 Sol cuts the cached input cost further — from $0.20 to $0.10 per million tokens — and positions itself explicitly as the near-Astra workhorse tier. The implied message: Astra is for use cases where you've proven you need it. Sol is for everything else.

The cost trajectory over three years:

The pattern is clear: the capability tier that cost $10 input in 2023 now costs $2 input in 2026. The frontier has moved upward, but near-frontier capability has dropped by 80% in three years.

What GPT-6.1 Sol Actually Is

The DevDay announcement described GPT-6.1 Sol as an upgrade to GPT-6 Sol with improved benchmark performance and a halved cached-input price. The key specifications:

• Pricing: $2.00 input / $10.00 output / $0.10 cached input per million tokens
• Long context (>272K input tokens): $4.00 input / $15.00 output per million tokens
• Batch and Flex: Half of standard rates — $1.00 input / $5.00 output
• Context window: Approximately 1.05–1.1 million tokens
• Benchmark: DeepSWE v1.1 at 75.2% (high effort, OpenAI-reported)
• Intelligence Index: Score of approximately 52, at $0.72 per index task
• Ultrafast version: Listed as coming soon; not available at launch

The benchmark figures require qualification. DeepSWE v1.1 scores are OpenAI's own reported evaluations — the same methodological issue that Signal's analysis of evaluation infrastructure identified as a systemic problem: model developers self-reporting on benchmarks they had access to during training introduces selection bias that independent evaluators consistently find inflates reported performance. The 75.2% figure should be treated as an upper bound on Sol's performance on that specific benchmark under optimal conditions.

The benchmark comparison analysis from Lindy found that GPT-6.1 Sol outperforms Anthropic's Claude Opus 5.5 at medium effort on AutomationBench at comparable or lower cost — a direct competitive positioning against Anthropic's second-tier model. The cross-provider comparison is the more practically relevant number for enterprise buyers than the absolute DeepSWE score.

The Commodity Threshold

The concept of commoditization in AI has been discussed since 2024, but most of the discussion has been vague about what commoditization actually means operationally. GPT-6.1 Sol's pricing makes it concrete.

Commoditization in AI occurs when near-equivalent performance is available from multiple providers at the same price point, and when the cost structure reaches a level where it stops being a constraint on product design. The first condition — multiple near-equivalent providers — has been true for most workloads since mid-2025. The second condition — cost below design constraint — is where GPT-6.1 Sol changes the picture.

Signal's earlier analysis of the AI native pricing crisis identified the threshold below which per-token inference cost stops constraining product architecture for most SaaS workloads. At $2 per million input tokens, a product running 500 input tokens per user interaction can handle 2 million interactions per million dollars of inference cost. For most enterprise SaaS products, that math means inference cost drops below 1% of revenue at reasonable pricing assumptions — the same economic zone that database and hosting costs occupy in traditional software. When infrastructure costs are below 1% of revenue, they stop driving architecture decisions.

GPT-6.1 Sol is not the first model to break this threshold — some use cases got there with earlier pricing. But it is the first model where OpenAI has explicitly said "this is the standard tier, the premium tier costs five times more, and the standard tier has near-equivalent performance." That explicit labeling matters because it creates a Schelling point: enterprises now have OpenAI's own authority for the claim that Sol is sufficient for most workloads, which makes it politically easier to deploy Sol rather than Astra in enterprise procurement conversations where legal and finance teams want vendor validation.

The Enterprise Budget Problem

The commoditization signal creates a specific problem for enterprises that built their 2026 AI budgets on Astra-tier pricing assumptions.

A Fortune 500 company that in Q2 2026 approved a $10 million annual AI inference budget based on routing its core workloads through Astra-tier models now has a question to answer: if Sol can handle 70-80% of those workloads at one-fifth the cost, is the $10 million budget still needed, or has the right budget become $2-3 million? The technology answer is probably the lower number. The organizational answer is more complicated.

Enterprise budget conversations about AI infrastructure tend to lock in assumptions about model tier and cost structure at the time of initial approval. Revisiting a budget mid-year — even to reduce it in response to better economics — requires a procurement reopening that most finance teams are reluctant to trigger. The result is that many enterprises will spend 2027 running Astra-tier workloads that could have been running on Sol, not because Sol is inadequate but because the budget architecture locked in before Sol existed.

The forward-looking implication for OpenAI's IPO position is worth noting: a permanently tiered pricing structure where the near-equivalent standard tier is 80% cheaper than the premium tier changes the revenue mix of the API business. OpenAI is betting that the standard tier generates enough volume to offset the per-token revenue reduction — the same bet AWS has made repeatedly with commodity compute pricing.

The Competitive Cluster

GPT-6.1 Sol's pricing puts it in a tight competitive cluster with several other near-frontier models that have converged in the $2-5 input per million range.

(Prices are approximate based on available third-party sources; verify against provider pricing pages before use in budget models.)

The cluster is significant for enterprise procurement: at these price points, model selection is no longer primarily a cost decision — it is a quality, latency, and vendor-relationship decision. An enterprise choosing between GPT-6.1 Sol and Gemini 4 Argon at identical input pricing is making a decision based on benchmark performance for their specific workloads, latency requirements, data handling policies, and existing vendor relationships. Cost has ceased to be the primary differentiator within this cluster.

This is what commoditization looks like in practice: not zero margin, not identical products, but a competitive cluster where the pricing range is narrow enough that non-price factors dominate the selection decision. The pricing analysis from szymonpaluch.com noted that Sol's pricing effectively forces Anthropic and Google to defend their positions in this cluster on quality and integration grounds, not on cost.

The Enterprise AI Buyer Playbook

The GPT-6.1 Sol launch changes the enterprise AI buying decision across three dimensions: model selection, budget architecture, and vendor relationship management.

1. Audit your current workload mix against the Sol/Astra split. The first action for any enterprise AI buyer is a workload audit: categorize every AI use case by output stakes and task complexity, and identify the fraction that requires Astra-tier versus Sol-tier capability. Most organizations will find that 60-80% of token volume can move to Sol without measurable quality degradation. The remaining 20-40% warrants Astra or independent evaluation. This audit is not a one-time exercise — as AI use cases expand, new workloads should be evaluated against the tier split before provisioning.

2. Run workload-specific A/B tests before committing to tier. The GPT-6.1 Sol benchmarks are OpenAI-reported and tested against specific evaluation sets that may not represent your workloads. Before committing a workload to Sol in production, run a controlled comparison against Astra on 1,000-5,000 real examples from your actual task distribution and measure quality differences that matter to your specific outcomes — not abstract benchmark scores. Most enterprise workloads will show no material quality difference. The ones that do are the candidates for Astra routing.

3. Rebuild your budget architecture around dynamic tier routing. The least efficient AI procurement model is static tier assignment: "all workloads use Astra" or "all workloads use Sol." The most cost-efficient model is dynamic routing: a classification layer that categorizes each incoming request by complexity and stakes, and routes to Sol by default with Astra reserved for requests that match a defined set of criteria (contract length above threshold, customer tier above threshold, explicit user upgrade request). This routing layer adds minimal latency — 5-10ms for a classification call — but can reduce inference costs by 60-75% for typical enterprise workload distributions.

4. Lock in Sol pricing in multi-year contracts before the competitive cluster stabilizes. The current window — a competitive cluster at $2-3 input per million with multiple providers offering near-equivalent capability — is the maximum bargaining leverage point for enterprise buyers. When the competitive cluster stabilizes and consolidates (as compute clusters always do), switching costs will increase and pricing power will shift back to providers. Enterprise AI teams with high token volume should use the current period to negotiate multi-year rate agreements at or near the current Sol pricing level, with volume commitment discounts, before the market settles.

5. Separate model selection from vendor selection in your architecture. The fastest-evolving variable in the AI cost landscape is which model is cheapest at a given quality level — a ranking that changes every few months as new model versions launch. Enterprise architectures that have hard-coded a specific model into their application layer pay a tax every time the model landscape shifts: migration costs, regression testing, and redeployment. Architectures that route through a model abstraction layer — where the model is a configuration parameter, not a code dependency — can take advantage of pricing shifts within hours of a new launch. The short-term investment in abstraction pays compounding dividends as the model market continues to evolve.

When Commodity Pricing Meets Commodity Quality

The commoditization of near-frontier AI creates a second-order problem that is less discussed than the pricing question: when multiple models offer comparable quality at comparable prices, what is the sustainable differentiation strategy for AI-powered products?

The answer that has emerged across the most durable AI-native products in 2026 is workflow integration and switching cost — exactly the pattern Signal identified in the analysis of the PLG ceiling. A product whose AI output is consumed inside a workflow — where it feeds downstream steps, where its outputs are stored in the product's own data layer, where users have built habits around reviewing and acting on its outputs — has switching cost that commodity model pricing cannot erode. A product that exposes AI capability as a standalone feature, with outputs that don't persist or integrate, has the same commodity problem as the underlying model: it can be replaced by any equivalent at any lower price.

The strategic implication for product teams is to build the workflow layer aggressively, while model costs are still low enough to subsidize it. At $2 input per million tokens, an enterprise SaaS product can run extensive AI processing on user data — enriching records, generating summaries, flagging anomalies, suggesting actions — without charging users explicitly for the compute. That "free" AI processing is what builds the data and workflow integration that creates switching cost. Products that run this playbook in 2026 will have structural moats by 2028 that are not replicable at any model price point.

Takeaway

Takeaway: GPT-6.1 Sol's one-fifth pricing relative to GPT-6 Astra is not a discount — it is OpenAI formally acknowledging that near-frontier intelligence has commoditized. The enterprise implication is a five-part playbook: audit workloads against the Sol/Astra split, run workload-specific A/B tests before committing to tier, rebuild budget architecture around dynamic routing, lock in Sol pricing in multi-year contracts before the cluster stabilizes, and separate model selection from vendor selection in application architecture. The deeper strategic signal is that AI pricing is approaching the zone where it stops constraining product design — the same economic zone that database costs reached a decade ago. The products that will win are not those with the best model access but those that use the low-cost window to build workflow integration deep enough that switching cost outlasts the pricing advantage.