OpenAI vs Anthropic Pricing 2026: The Replit Deal
OpenAI vs Anthropic pricing is now negotiable: Replit reportedly moved a $100M+ contract on a 30%+ discount. What the deal proves, and how buyers can ask too.
Key Takeaways
- Replit reportedly moved a contract worth over $100 million a year from Anthropic to OpenAI after OpenAI offered a discount of more than 30% — and Anthropic declined to extend Replit's previous discount.
- The reported trigger was price, not a benchmark: no capability test appears anywhere in the reporting of the switch.
- The split is messier than the headline: Replit later said it remains model agnostic, with OpenAI as the default for Replit Agent and Claude still selectable.
- On published list prices the two flagships now cost the same — GPT-5.6 Sol at $4/$20 and Claude Opus 5.5 at $4/$20 per million tokens — so the real price for big buyers lives in negotiated discounts.
- Smaller buyers can use the same levers at their own scale: competing quotes, committed usage, batch processing, cheaper models for routine work, and a credible fallback option.
OpenAI vs Anthropic Pricing 2026: The Replit Deal
If you pay for AI, here is the business story of the month: a contract worth more than $100 million a year reportedly moved vendors on price alone. Not on a benchmark. Not on a new feature. On a discount.
According to The Information's reporting, summarised by AI Tools Recap on October 25, 2026, OpenAI offered Replit a discount of more than 30% on a deal worth over $100 million a year — after Anthropic declined to extend Replit's earlier discount unless it committed to spending more. Replit then made OpenAI the default model provider for Replit Agent.
That is the first clear, reported evidence that frontier AI pricing — what developers actually pay, not the published page — is negotiable, even at the very top of the market. If a $100 million account can move on price, a $100,000 one can at least ask.
One caveat first: the discount terms are reported, not confirmed by either vendor in public. Everything below about the deal's terms is attributed reporting; the list prices are verified from each vendor's own pricing page.
What actually happened
- October 16, 2026: Replit chief executive Amjad Masad told TechCrunch Disrupt in San Francisco that Replit was no longer working with Anthropic, citing "cost, partnership, and performance" — and adding, "I think ChatGPT is better now."
- October 24, 2026: The Information reported the price: more than 30% off a contract worth over $100 million a year, after Anthropic ended Replit's earlier discount.
- Days later: the picture softened. Replit clarified it "remains model agnostic" and that OpenAI is now its preferred partner for Replit Agent. TechCrunch was told the Anthropic partnership was still in place, and Replit's documentation still listed Claude models.
That messiness matters. In platform businesses, the default is where the money is: most users never change it. OpenAI becoming the default captures the bulk of Replit's AI spending even if Claude remains available.
The interesting absence: no benchmark anywhere
No benchmark appears anywhere in this story. Replit builds a coding product — the use case where model differences should matter most. If the switch had been about Claude losing on code, that would have been the headline. Instead, the reported trigger is a withdrawn discount. Once a model is wrapped in a company's own harness, with retries and scaffolding around it, the two leading options were apparently close enough that a 30% price difference outweighed whatever capability gap remained.
It is not just Replit: the wider price war
The Information's reporting describes a broader OpenAI discount campaign aimed at Anthropic's large customers, with discounts reported as high as 90% in some cases. Rippling reportedly shifted its coding workload from 16% OpenAI / 57% Anthropic in May to 52% / 18% now. DataCamp's chief executive reportedly said OpenAI's pricing cut his company's AI costs by 80–90% — a reported executive claim, not an audited saving. And a US bank with over $7 billion in assets reportedly had its ChatGPT seats doubled to 500 at no extra charge, bringing its per-user cost under $7 against a standard $20-plus seat price.
These are reported figures, not vendor confirmations, but the direction is consistent across multiple companies, and it matches the timing pressure on Anthropic: its first profitable month was reported in August, and it filed its S-1 for an IPO in October (see our breakdown of Anthropic's S-1 filing and its $80 billion loss).
List prices vs. real prices: the comparison table
On published list prices, the two flagships cost exactly the same. Rates verified from the vendors' own pricing pages as of October 25, 2026 — check the live pages before budgeting.
| Model | Input / 1M tokens | Output / 1M tokens | Notes | | --- | --- | --- | --- | | GPT-5.6 Sol (promo) | $4 | $20 | Promo through at least Nov 21, 2026 (standard $5/$30) | | Claude Opus 5.5 | $4 | $20 | Batch API halves both to $2/$10 | | GPT-5.6 Terra | $2 | $12 | Mid-tier OpenAI model | | Claude Sonnet 5.5 | $2 | $10 | Batch: $1/$5 | | GPT-5.6 Luna | $0.20 | $1.20 | Cheapest OpenAI tier | | Kimi K3.1 Pro | $2 | $10 | Open weights promised by Nov 4, 2026 |
Two lessons follow. First, the headline rate is a poor predictor of the bill: caching, batch processing, long-context surcharges (OpenAI charges more above ~272,000 input tokens; Opus 5.5 adds none) and actual token usage matter more than a dollar or two per million. Second, the open-weight tier is about to argue from below: Kimi K3.1 Pro scores 55 on Artificial Analysis's Intelligence Index against Opus 5.5's 58, at half the flagship list price — and if its weights ship, some buyers can run it themselves. See Kimi K3.1 Pro explained.
How a smaller buyer can actually negotiate
You will not get Replit's deal by asking nicely. But the same levers work at smaller scale, and they cost nothing to try:
- Get a competing quote in writing. Run the same workload on both APIs for a week — your own numbers beat any benchmark.
- Negotiate on committed usage, not list price. Discounts usually trade a lower rate for a spending commitment or longer term. Only commit to volume you already use.
- Ask about batch and caching explicitly. Batch processing halves Anthropic's rates, and prompt caching cuts input costs sharply on repeated context — no negotiation needed.
- Route routine work to cheaper models. Sending classification, extraction and drafting to a Sonnet-, Terra- or Luna-class model often saves more than any discount.
- Negotiate before renewal, in writing. Ask what the rate becomes at expiry, whether it needs a minimum spend, and what happens to your data if you leave. Keep a fallback working: if switching takes a quarter of engineering, you have no leverage; if it takes a config change, you do.
None of this guarantees a discount — smaller accounts typically get published levers rather than bespoke rates. But the Replit story proves the list price is not the final word.
When you do not need to negotiate at all
Not every buyer needs an enterprise conversation. Small usage is covered by the free tiers of ChatGPT and Claude, and pay-as-you-go API billing means a side project can cost a few dollars a month. For occasional work, a $20-a-month subscription tier is cheaper than any API negotiation. And if your workload is open-ended, the free and open-source AI tools we rounded up this month — including models you can run yourself — remove the vendor from the pricing question entirely.
The right question is not "which vendor is cheapest?" but "what does a completed task cost me?" Measure cost per finished job, including retries and review time.
What the Replit deal does not prove
- It does not prove OpenAI is better at coding. Masad's performance comment is one chief executive's opinion; the reporting attributes the switch to price.
- It does not prove Anthropic is in trouble. Losing a default at one large customer while filing for an IPO is awkward, not existential — and its reported first profitable month in August cuts against any simple decline story.
- It does not prove your discount will last. Promotional pricing ends and negotiated rates reset at renewal. Price the renewal, not just the first year.
The bottom line
AI prices are now negotiable in a way they were not a year ago, because the leading models are close enough that buyers can credibly walk away — and because the open-weight tier led by Kimi is about to give them somewhere to walk to. For anyone choosing between ChatGPT, Gemini and Claude in 2026, the rule is: check the list price, then what the vendor will actually charge you, then what it costs to leave. The cheapest model is the one you can switch away from.
No affiliate links in this article — tool names are plain text, and all prices above are from the vendors' own pricing pages or clearly attributed reporting. Featured image: an original AI-generated illustration, not a photograph.
Sources
FAQs
Can a small company really negotiate AI pricing, or is that only for $100M accounts?
Published enterprise discounts are negotiated case by case, and no vendor publishes a small-buyer discount menu — so nobody can promise you a percentage. What the Replit reporting proves is that list prices are a starting point, not a final price, at the top end. At smaller scale, the realistic levers are a competing quote, a usage commitment, batch or off-peak processing, and routing routine work to a cheaper model. Ask in writing before renewal, when your leverage is highest.
Is OpenAI now cheaper than Claude?
On published list prices, no — the flagships are tied. GPT-5.6 Sol's promotional price is $4 per million input tokens and $20 per million output tokens, and Claude Opus 5.5 lists at exactly $4 and $20. What differs is the negotiated enterprise price, which is private, and the surrounding costs: caching, long-context rules, batch discounts and how many tokens your workload actually uses.
Did Replit stop using Anthropic completely?
Not cleanly. Replit's chief executive first said on stage that the company was no longer working with Anthropic, then clarified that Replit remains model agnostic, with OpenAI as its preferred partner and the default for Replit Agent. Reporting after the announcement said the Anthropic partnership was still in place and Replit's documentation still listed Claude models. The default changed; Claude did not vanish.
Will an AI discount like Replit's last?
That is the open question. Promotional and negotiated prices can expire, reset at renewal, or depend on spending commitments — GPT-5.6 Sol's promotional pricing, for example, is stated to run at least through November 21, 2026. Before switching for price alone, ask what happens at renewal, whether the discount requires a minimum spend, and what it costs to switch back.
When does Kimi K3.1 Pro's open-weight release change this negotiation?
Moonshot AI has promised to publish Kimi K3.1 Pro's full model weights by November 4, 2026, with API pricing already at $2 input and $10 output per million tokens. If the weights ship with usable licence terms, buyers gain a self-hosting fallback that strengthens any negotiation — but until the licence is public, treat it as a promising lever, not a done deal.
Should I switch AI vendors just because of a discount?
Only if the cheaper option passes your own tests on your own workload. A discount on tokens you burn through retries, or on a model that needs more human review, can cost more than it saves. Run the same tasks on both vendors, measure cost per completed task — not per token — and keep a fallback model working before you move production traffic.
Written by
AI Trends Daily Team
The AI Trends Daily team researches AI and technology daily, then explains it in plain English — no jargon, no hype, just what matters.
Related Articles
AIKimi K3.1 Pro Explained: Open-Weight AI Nears Opus 5.5
Moonshot AI's Kimi K3.1 Pro launched Oct 23, 2026, scoring 65.7% on Terminal-Bench 4.0 — one point behind Claude Opus 5.5. Open weights promised by Nov 4.
October 24, 2026
AIClaude Duet vs Meta Muse vs OpenAI Dots: 2026 Review
Claude Duet vs Meta Muse vs OpenAI Dots: an honest 2026 comparison of pricing, availability and safety controls — plus whether you even need an always-on AI agent.
October 23, 2026
AIRecursive Flint: The Free Coding Model Built by AI (2026)
Recursive Flint is a free 40B coding model that runs offline on your laptop — built by AI in 4 months. See how it compares to Cursor, Copilot and Claude Code.
October 22, 2026