🔍 Read the full analysis: OpenAI Vs. Anthropic Pricing: What To Check Before You Commit on ThorstenMeyerAI.com
Get business pricing on office and shipping supplies
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
OpenAI and Anthropic have cut prices for several models, while reporting describes OpenAI offering discounts to attract enterprise customers. The headline claim of 80% to 90% savings refers to one buyer’s expected total cost reduction, not a general discount off OpenAI’s list prices. Buyers should compare workload-specific costs, usage limits, credits and contract terms before committing.
OpenAI and Anthropic have cut prices on several AI models in recent months, and OpenAI is also offering discounts to some large customers, according to reporting by The Information. The changes give companies more room to negotiate, but list prices alone do not show what a buyer will pay: usage allowances, expiring credits, seat minimums and workload-specific costs can materially change the bill.
Pricing figures in the source report, checked against vendor announcements and pricing pages on October 26, show OpenAI’s Sol model moving from $5 per million input tokens and $30 per million output tokens in June to $2 and $10 for GPT-6.1 Sol in September. Its Luna model moved from $1 and $6 to $0.10 and $0.50. These are prices for successive models, not necessarily reductions to the price of an unchanged model.
Anthropic also lowered some costs. Its Opus line went from $5 per million input tokens and $25 per million output tokens for Opus 5 to $4 and $20 for Opus 5.5; cache reads also became cheaper. Anthropic has said Opus 5.5 costs 40% less on typical workloads than Opus 5, and that its Fable 5.1 model can cost less on agentic tasks. Those workload savings are the company’s estimates, not a guaranteed result for every customer.
The Information reported on October 24 that OpenAI was offering discounts to win customers. One reported Replit deal involved a discount of more than 30% on a contract valued at over $100 million a year. DataCamp CEO Jonathan Cornelissen described an expected 80% to 90% reduction in his company’s total AI costs. That figure is not a discount of that size from OpenAI’s list price. Separately, Rippling CEO Parker Conrad said OpenAI’s share of his company’s coding work rose from 16% in May to 52%, while Anthropic’s share fell from 57% to 18%; he attributed the shift to cost.
Infographic 01 · The price slide
Four months, a fraction of the price.
What a million output tokens cost on each vendor’s price list, from the June previews to today. Bars show the price; shorter is cheaper.
OpenAI · output price per 1M tokens
List prices, standard tier
Anthropic · output price per 1M tokens
List prices, standard tier
Same scale in both panels. Fable 5.1 kept its $50 output price and cut cache reads instead.
The quieter cut: cache reads
matters most for agents
Cache-read price per 1M tokens, same scale in all three.
Woolf estimates cached tokens make up 96–99% of token spend in agentic coding. Anthropic says Opus 5.5 costs 40% less than Opus 5 and Fable 5.1 about 25% (up to about 45%) less than Fable 5 on typical workloads; those are vendor claims. Most rows compare a successor model with its predecessor. Anthropic’s newer tokenizer produces about 30% more tokens for the same text, so per-token prices are not directly comparable across vendors.
Sources: OpenAI API pricing (26 Oct 2026); OpenAI announcements (26 Jun–29 Sep 2026); Anthropic API pricing (26 Oct 2026); Anthropic announcements (30 Jun–22 Sep 2026); Vercel AI Gateway index (Oct 2026); Max Woolf (25 Oct 2026).
Powered by Thorsten Meyer AI
How Contract Terms Change AI Costs
Falling token prices may reduce the cost of using AI, but a company’s actual bill depends on how much it uses each model, which tasks it assigns, and what its contract includes. For coding agents, repeated reading of cached conversation context can make cache pricing a major part of the expense. A model with a higher headline rate may still cost less on a particular workload, while a lower rate may not help if the model needs more tokens or performs worse on the buyer’s tasks.
Plan terms can also change the value of a deal. The source report says Anthropic added capped, expiring API credits to its Max and Team plans and that its Enterprise plan requires 20 seats. It also says OpenAI reduced usage included with new $200 Pro subscriptions without changing the price. Buyers comparing offers should check the amount of included usage, what happens when it runs out, seat commitments, credit expiration and whether a discount lasts beyond the initial contract period.
The competition creates negotiating leverage, especially for organizations able to move work between providers. But switching has costs of its own: teams may need to change software, review privacy and security arrangements, and test whether a replacement model handles their tasks reliably. A lower quote is not the same as a lower total cost unless the buyer compares equivalent workloads and service terms.
Products Worth Considering
Dell UltraSharp 27 Thunderbolt Hub Monitor (U2726DER)
Samsung 27 Odyssey OLED G5 (G50SF) QHD QD-OLED Gaming Monitor
LG 34GX90SA-W 34-inch Ultragear WQHD OLED Curved Gaming Monitor
As an affiliate, we earn on qualifying purchases.
Recent Model and Price Changes
The reported reductions followed a series of model launches and price changes from June through September 2026. OpenAI cut Luna pricing in July, temporarily reduced Sol pricing in August and introduced GPT-6 models at lower prices in September. Anthropic made an introductory Sonnet price permanent, released Opus 5.5 and lowered cache-read prices for another model. Data scientist Max Woolf tracked the sequence in an October 25 post; the source article says its figures matched vendor announcements and pricing pages checked the next day.
These comparisons need care because many lower prices apply to new model versions rather than the same model at a reduced rate. Vendors have also made claims about performance and savings on typical or agentic workloads. Buyers need to test those claims against their own tasks, since results depend on model behavior, prompt and cache patterns, and the amount of human review required.
Enterprise switching is already part of the story, though the public examples are limited. The Information’s reporting described OpenAI discounts aimed at major customers. Replit CEO Amjad Masad said at an October 16 event that his company was no longer working with Anthropic and that he thought ChatGPT was better; he later told TechCrunch that Replit remained model-agnostic and OpenAI was its preferred partner. TechCrunch also reported that a person familiar with the partnership said it remained in place.
“OpenAI’s share of coding work rose from 16% in May to 52%, while Anthropic’s fell from 57% to 18%.”
— Rippling CEO Parker Conrad, as reported by The Information
What Buyers Still Need to Verify
The public reporting does not provide complete contracts or enough detail to compare every discount. It is unclear how long the reported enterprise offers last, which usage thresholds or service conditions apply, and whether other customers can obtain similar terms. The Replit accounts also leave some uncertainty about the state of its partnership: Masad described OpenAI as the preferred partner, while a source cited by TechCrunch said the partnership remained in place.
It is also not clear whether quoted savings include all costs, such as model testing, software integration, employee use, support and work assigned to other vendors. The 80% to 90% DataCamp estimate is a company-specific expectation, not an independently verified saving or a general forecast. Model benchmarks and vendor estimates cannot establish how a model will perform on a particular organization’s data and tasks.
Several plan and pricing details may change. Buyers should confirm current rates directly with vendors and check whether prices apply to input and output tokens, cached tokens, batch processing or other usage categories. They should also verify what counts toward a plan’s allowance and whether usage is capped, reset or billed separately.
Steps Before Signing a Deal
Before renewing or selecting a provider, buyers can run the same representative tasks through both services and record cost per completed task, not just token price. The test should include typical and demanding cases, cached and uncached usage, output quality, latency, error rates and the amount of human correction needed. Results should be evaluated against the organization’s security and data-handling requirements as well as its budget.
Procurement teams should ask each vendor for a written breakdown of rates, discounts, seat minimums, usage allowances, credit expiration, overage charges and renewal terms. They can also ask whether discounts are temporary, conditional on volume or limited to specific models. If a company can route work across providers, it may preserve flexibility, but should include the operational cost of maintaining that arrangement in its comparison.
The next meaningful comparison will come from current vendor quotes and measured customer workloads, not a single headline discount. Pricing and plan terms remain subject to change, so organizations should check them again before signing and monitor usage after deployment.
Key Questions
Does the 80% to 90% figure mean OpenAI is offering that discount to all customers?
No. The figure refers to DataCamp CEO Jonathan Cornelissen’s estimate of how much his company’s total costs would fall under OpenAI’s offer. It is not a general discount from OpenAI’s list prices.
Which vendor has the lower token prices?
It depends on the model and token type. The source report lists lower prices for some newer OpenAI models, while Anthropic also cut prices and cache costs. Compare current input, output and cached-token rates for the exact models and workload you plan to use.
Why do cache prices matter for AI coding tools?
Coding agents may reuse and reread conversation context as they work. If a large share of a workload uses cached tokens, the cache-read rate can have a substantial effect on cost. The share varies by tool and use case, so buyers should measure it in their own tests.
What should companies check besides the listed price?
Check included usage, seat minimums, expiring credits, overage charges, discount duration and renewal terms. Also test output quality, latency, security requirements and the staff time needed to review or correct results.
Source: ThorstenMeyerAI.com
