AI Price Cuts Are Real, but Enterprise AI Bills May Not Shrink With Them

Week of October 5, 2026

Google Halves Image Prices With Nano Banana 2.1

Google's new fast image model cuts the cost of 1K and 4K images by about half, one of three AI price cuts this week.

• Previous price
• Nano Banana 2.1

View the data

	Previous price	Nano Banana 2.1
1K image	6.7¢	3.36¢
4K image	15.1¢	7.56¢

More charts in By the Numbers

Three of the biggest AI providers cut prices this week. Anthropic released Claude Haiku 5.5 at about a quarter of the average cost of its predecessor, OpenAI opened a new API that bills only for input, and Google halved the price of its fast image model. For CIOs and CFOs watching AI spending climb, that sounds like relief. The fine print, and the way enterprise AI is actually being used, suggests the per-token price is only one part of the bill, and not always the part that decides it.

How big are this week's AI price cuts?

Anthropic launched Haiku 5.5 on Wednesday, the third model in its Claude 5.5 family in about a month. For prompts under 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, and Anthropic says it is about 75% cheaper on average than Haiku 4.5, Reuters reported. The company is aiming it at classification, summarization, extraction, live support and voice agents, the kind of high-volume work that makes up much of enterprise AI traffic.

Early in the week, OpenAI released its Decisions API in public beta. It returns structured answers, such as a yes-or-no probability, a choice from a fixed list or a severity score, for tasks like routing complaints or deciding an agent's next step. It runs on GPT-6 Luna and charges $0.10 per million input tokens, with output and caching not billed, Unite.AI reported. OpenAI's API changelog says it answers 10 times faster than its general Responses API, a claim that has not been independently tested.

Google, meanwhile, released Nano Banana 2.1, which cuts the price of a 1K image from 6.7 cents to 3.36 cents and a 4K image from 15.1 cents to 7.56 cents, The Decoder reported. The model is rolling out across Google products including the Gemini Enterprise platform.

Why might enterprise AI bills not fall as fast as prices?

The first reason is in the footnotes. Anthropic's headline cut is 90% for shorter prompts, but its average saving is about 75%, partly because Haiku 5.5 uses a new tokenizer. Anthropic's developer documentation puts the increase at roughly 30% more tokens for the same text, Mixed reported, and it advises developers to recount prompts and revisit cost estimates. Prompts above 100,000 tokens are billed at five times the lower rate, which matters for agents that carry long histories or large documents.

The second reason is volume. Agents do not make one call per task; they plan, call tools, check results and retry. Cheaper tokens tend to invite more of them. Google seemed to acknowledge the problem when it launched its new enterprise Gemini agent on Thursday with smart model routing and real-time spend caps, TechCrunch reported. Futurum Group analyst Keith Kirkpatrick said those controls target CFO worries about unpredictable agent costs, and that the cost question now turns on how Google meters agent activity, with pricing details still to come. "The open question is commercial," he wrote.

The third reason is that quality claims are still vendor claims. Anthropic published large benchmark gains over Haiku 4.5, and OpenAI published speed figures. Until enterprises test them on their own workloads, the real cost per completed task is unknown.

Will low AI prices last?

There are two reasonable readings of this week. One is that competition is doing what it should. Three labs shipped cheaper options within days of each other, and enterprises with clear, repetitive tasks can now run them at a fraction of last year's cost.

The other reading is more cautious. Reuters noted that Anthropic's launch came ahead of a planned IPO, and the week brought fresh questions about AI lab finances. OpenAI reportedly told investors its annualized revenue is approaching $50 billion, about $20 billion less than earlier reports, TechCrunch reported, citing the Financial Times. The gap appears tied to differences in how OpenAI and Anthropic count revenue sold through cloud partners, so it does not necessarily mean revenue fell. But it is a reminder that today's prices are set by companies spending heavily to win market share.

Louis Dimodugno, Global Chief Data Officer at Verisk, made that case at The AI Enterprise Conference 2026. He said token prices were changing almost daily as frontier providers competed, and that prices were likely to rise once those companies go public, catching organizations without visibility into their AI use off guard. "The financial piece is the piece that's always going to get the most attention," he said, in a session on how Verisk lets AI savings fund new AI work.

Both readings can be true. Prices for small, fast models are falling now, while the long-term price of frontier capability is uncertain.

What should enterprise leaders do with cheaper models?

The practical response is to treat this week as a reason to re-architect, not just to re-budget.

• Right-size by task. "You don't need premium models for everything that you're doing," Pushkar Rege, Vice President and Co-Founder of the Business Intelligence Program at Mastercard, said at The AI Enterprise Conference 2026, in a panel on measuring AI cost per outcome. Classification, routing and extraction are exactly the jobs this week's cheaper models target.
• Measure cost per task, not per token. Tokenizer changes, context tiers and agent retries all shift the real cost. Rerun your own evaluations and compare the full cost of a completed task before switching models.
• Build routing and caps into the platform. Model routing, per-project budgets and chargeback by department are becoming standard features. If your vendors do not offer them, build them into a shared layer. Speakers at a AI Connect panel on AI infrastructure and inference costs added related levers: run only inference on GPUs, match infrastructure to each workload's service level, and cache repetitive queries.
• Negotiate for the long term. If prices may rise after IPOs, multiyear commitments with price protections, and architectures that can switch providers, are worth more now than the last few cents per million tokens.

The cheapest token is the one an enterprise never needed to send. The companies that benefit most from this week's cuts will be the ones that already know what each AI task is worth to them.

Frequently asked questions

How much does Claude Haiku 5.5 cost?

For prompts under 100,000 tokens, $0.10 per million input tokens and $0.50 per million output tokens. Longer prompts cost $0.50 and $2.50. Anthropic says it is about 75% cheaper on average than Haiku 4.5.

What is OpenAI's Decisions API?

A public-beta API on GPT-6 Luna that returns typed answers such as probabilities, choices or severity scores for tasks like routing and classification. It bills $0.10 per million input tokens and does not charge for output.

Will AI model prices keep falling?

Prices for small, fast models are falling now amid competition. Some enterprise data leaders expect prices to rise once AI labs go public, so cost visibility and flexible architectures matter.

Disclosure: AI Enterprise Magazine uses AI tools in its editorial process, including Anthropic's Claude models. Anthropic had no role in this article.

Filed under: AI Strategy & ROI, Data & AI Infrastructure