AI Got Up to 80% Cheaper This Week. Your Bill Didn’t

On October 26, 2026, the price of artificial intelligence fell off a cliff — and if you buy AI through an app, an agent builder, or a monthly subscription, you probably haven’t seen a cent of it. OpenAI launched GPT-6.1 Terra at half the price of its predecessor the same week Anthropic’s IPO filing revealed a Claude price ladder that sits a fraction of where it stood in July. Headline API prices are down by a third to three quarters depending on the model line, and cached-input prices — the part that matters most for agent workloads — fell as much as 80%.

Here’s the short answer: the AI price war is real, it’s dated, and it’s measurable — but most small buyers rent their AI through middlemen who pocket the difference, or through subscriptions whose prices never move when API costs fall. If you want the savings, you have to go get them: renegotiate, route work by cost per task instead of cost per token, and cache aggressively. This article gives you the dated price table, the math that shows why your bill didn’t move, and a playbook for capturing the cut.

The dated price table: July vs October 2026

The dated price trail below comes from independent trackers that read the vendors’ own pricing pages, including Tao of Mac’s price-of-a-token log and MarkTechPost’s launch breakdown — use those pages to double-check any figure before you budget on it, since vendors revise prices without announcement. OpenAI’s own GPT-6.1 Terra model page lists the $1 input, $0.05 cached input, and $5 output rates directly.

Two things jump out. First, the cuts are broad: every major line got cheaper, not just one loss-leader model. Second, the cache column fell further than the headline column almost everywhere. Terra’s cache reads went from $0.25 to $0.05 per million tokens — an 80% cut, and the source of the “up to 80%” headline. When you see that figure quoted, it’s the cache number doing the work, not the headline rate. Both are real; they just measure different things.

The trigger event was OpenAI’s October 26 launch of GPT-6.1 Terra: $1 per million input tokens, $0.05 cached input, $5 per million output — half the price of GPT-6.1 Sol and a quarter of Claude Opus 5.5’s list price. Terra is a mid-tier model built for computer use: reading screens, filling forms, drafting proposals, editing video through apps. It has a roughly 1.05-million-token context window and up to 128,000 output tokens. It’s available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu subscribers — and this is the trap most readers will hit — it is not in standard ChatGPT chat. Free and Go tiers aren’t listed. It’s also available via API, Amazon Bedrock, and GitHub Copilot.

Anthropic’s side of the story arrived the same day, not as a launch but as a filing: its S-1 for a Nasdaq listing revealed the price ladder its business now runs on, with Opus 5.5 at $4/$20 and Sonnet 5.5 at $2/$10. The filing also showed why both vendors can afford to cut: Anthropic reported $16.3 billion in first-half 2026 revenue, up from $1.16 billion a year earlier, and its first profitable quarter. Price wars are easy to fund when revenue is growing fourteenfold.

Why your bill didn’t fall: the three middlemen

If the price of the raw material fell 60%, why is your monthly AI spend the same? Because most small businesses don’t buy raw API tokens. They buy one of three things, and each one absorbs the cut differently.

1. Subscriptions. ChatGPT Plus, Claude Pro, and their business tiers charge a flat monthly fee. When the vendor’s per-token cost falls, your subscription price doesn’t — the vendor keeps the margin, or spends it on more features. Terra’s arrival in ChatGPT Work and Codex is a good example: OpenAI says the same plan covers more tasks with Terra than with its predecessor, which is a real benefit, but it’s not a price cut on your bill. You get more work per dollar only if you actually do more work.

2. Agent builders and apps. Tools like Copilot Studio ($200/month per 25,000 credits in recent comparisons), Lindy, n8n, and the rest buy API tokens wholesale and sell you credits, seats, or executions. Their prices are set by their own packaging, not by the API price list. When Terra halves the cost of the underlying tokens, a credit-metered builder can either pass the savings through, keep its credit prices and widen its margin, or quietly let you do more work per credit. Most do the second or third, and none of them send you a letter about it. This is also why credit-metered builders get expensive unpredictably: your bill depends on their credit math, not on the API price you read about in the news.

3. Enterprise-style negotiated deals — which you don’t have. The biggest discounts in this market aren’t on any price page. Reports put enterprise negotiated discounts as high as 90% off list, and Replit’s reported $100-million-plus switch from Anthropic to OpenAI on a 30%-plus discount shows how the game is played at scale: vendors will cut deep to win or keep a marquee account. A small business buying list-price API access or a $20 seat will never see those numbers — but knowing they exist tells you the list price is a starting point for negotiation, not a law of nature, even at modest volumes.

There’s a fourth reason, and it’s the most boring: most buyers never re-check. API prices changed four times in some model lines between July and October. If you set up an agent workflow in July, priced it once, and never looked again, you’re budgeting on stale numbers. The price war rewards people who re-price quarterly. Everyone else donates the difference.

The number that actually matters: cost per task, not cost per token

Here’s where the conversation usually goes wrong. People compare $1 vs $2 per million input tokens and conclude the cheaper model is “half price.” But nobody buys a million tokens; they buy completed work — an invoice processed, a form filled, a support ticket resolved. And on that measure, the gap between models is far bigger than the token price suggests.

Independent testing by Artificial Analysis puts GPT-6.1 Terra at $0.40 per test task on its Intelligence Index, versus $0.72 for GPT-6.1 Sol and more than $5 per task for Claude Sonnet 5.5 and Opus 5.5 at their highest settings. Terra scores 48.1 on the same index — below Sol (52), Sonnet 5.5 (56), and Opus 5.5 (58), but at a fraction of the per-task cost. OpenAI’s own launch claims put Terra 15 points ahead of its predecessor on an agentic benchmark at roughly a third of the cost per task, and its hallucination rate down from 53.6% to 16.7% on internal evals at low effort. Treat vendor benchmarks as vendor claims — they’re designed to flatter — but the independent direction agrees: Terra is the cheapest per task, not the smartest model, and that’s exactly the point for routine work.

Let me make it concrete with a fixed-volume example. Take a small job: 10,000 input tokens in, 2,000 output tokens out.

At this volume, the difference between Terra and Opus is six cents per job. Run that job 10,000 times a month and it’s $200 versus $800 — and that’s before the real multiplier kicks in.

The real multiplier: caching

The most underreported part of this price war is the cache column, and it’s where agentic workloads actually save money. An AI agent doing office work re-reads the same context over and over: the same invoice template, the same company handbook, the same customer history, the same long document. Cached input — tokens the model has already seen and stored — now costs a fraction of fresh input. Terra’s cache reads at $0.05 per million tokens are one-twentieth of its fresh-input price.

Think about what that means for a nightly invoice-processing job. The first invoice of the night pays full input price for the template and instructions. Every invoice after that reuses the cached context and pays almost nothing for it. The vendors cut cache prices hardest precisely because agents are their growth engine: cheap caching makes long-running agents affordable, and affordable agents lock you into their ecosystem. For you, the practical lesson is simple: if your agent tool supports prompt caching, turn it on and structure your prompts so the stable parts (instructions, templates, reference documents) come first and stay identical across runs. That’s free money the price war just made freer.

One honest caveat before you celebrate: reasoning tokens count. For reasoning models, the internal thinking tokens you never see are billed as output tokens, which are the expensive kind. A model that “thinks” for 8,000 tokens before writing a 500-token answer costs you for 8,500 output tokens. That’s part of why per-task costs for top models look so much worse than their token prices suggest, and it’s why routing decisions should be made on measured per-task cost, not on the price page.

The small-buyer playbook: capturing the cut

Here’s what I’d do this week if I were running a small business on AI tools.

1. Re-price everything you run. List every AI subscription, agent builder, and API bill. For each one, check the current list price of the underlying model against what you set up. If you built a workflow in July on GPT-5.6 Terra or Claude Opus 4 pricing, you’re almost certainly overpaying somewhere. This one audit is the highest-return hour in this article.

2. Route by cost per task, not by brand loyalty. Split your workloads into three tiers. Bulk, routine work (sorting, extracting, drafting first passes) goes to the cheapest competent model — right now that’s Terra-class pricing, or even budget lines like GPT-6 Luna and Claude Haiku 5.5 at $0.10/$0.50. Standard work goes to the mid-tier. Only brand-critical, high-stakes work (a legal letter, a final customer-facing answer, a complex analysis) earns the flagship price. Most businesses run everything on the flagship by default. That’s like hiring a senior partner to file your invoices.

3. Renegotiate or dual-vendor. If you buy API access directly and your volume has grown, ask for a discount — the market just gave you leverage, and vendors are demonstrably willing to cut for volume. If you buy through a builder, ask whether Terra-class pricing or cache discounts are passed through, and what your per-task cost actually is. If they can’t tell you, that’s information too. Running two vendors in parallel — even as a backup — keeps both honest and protects you when one changes prices or policies.

4. Turn on caching and stable prompts. As above: put stable context first, keep it identical across runs, and make sure your tool actually caches. Then measure: compare a week’s cost before and after. If caching doesn’t move your bill, your workload may not repeat context enough for it to matter — useful to know, and it changes which model tier you should buy.

5. Watch the trap in the other direction. Cheaper models invite more usage. An agent that cost $800 a month at old prices and $200 at new prices will tempt you to run four times as much work — and suddenly you’re back at $800, just busier. Set a monthly cap per workflow before you expand it, and treat “the model got cheaper” as a savings event first and an expansion opportunity second.

What to watch next

Three things will decide whether this price war keeps going or stalls. First, Anthropic’s IPO roadshow: the filing’s $265.6 billion in non-cancelable purchase commitments means both vendors are locked into enormous compute spending, which pressures them to keep volume growing — and price cuts are how you grow volume. Second, whether agent builders pass savings through or pocket them; the first builder to advertise “Terra pricing, passed through in full” will force the rest to answer. Third, the open-weight wildcard: independent tests put open-weight models like MiMo-V2.6-Pro at 46.3 on the Intelligence Index for $0.13 per task — cheaper than Terra per task, if you can host it. For most small businesses that’s still too much infrastructure, but the direction is clear: the floor keeps dropping.

The price war is real, the table above proves it, and the savings are available to anyone willing to re-price, re-route, and renegotiate. Your bill didn’t fall on its own because nobody in this market lowers your price out of generosity. Go collect it.

If you’re still choosing between the vendors rather than the prices, our earlier breakdown of Claude Haiku 5.5’s price cut shows how the budget tier changed the math for small users, and our look at DeepSeek and GLM’s October price moves covers the other end of the market — the ultra-cheap models that started this whole race.

Frequently Asked Questions

How much did AI API prices actually fall in October 2026?

Headline list prices fell by roughly a third to three quarters depending on the model line between July and October 2026. Claude Opus fell from $15/$75 to $4/$20 per million input/output tokens; OpenAI Terra fell from $2.50/$15 to $1/$5. Cached-input prices fell further — Terra’s cache reads dropped 80%, from $0.25 to $0.05 per million tokens.

What is GPT-6.1 Terra and where can I use it?

GPT-6.1 Terra is OpenAI’s fast, efficient mid-tier model launched October 26, 2026, built for computer use — reading screens, filling forms, drafting documents, and operating software. It’s available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu subscribers, and via API, Amazon Bedrock, and GitHub Copilot. It is not in standard ChatGPT chat.

Why hasn’t my ChatGPT or Claude subscription gotten cheaper?

Subscriptions charge a flat monthly fee that doesn’t track API token prices. When API costs fall, vendors typically give you more work per plan or keep the margin, rather than cutting your monthly price. Your bill only falls if you buy API access directly and re-price, or if you switch to a cheaper tier.

What is cost per task, and why is it better than cost per token?

Cost per task is what you actually pay to get one job done — including reasoning tokens, retries, and overhead. Independent testing put GPT-6.1 Terra at about $0.40 per test task versus more than $5 for top-tier models at their highest settings, a far bigger gap than the token prices suggest. Routing work by measured per-task cost saves more than comparing price pages.

Should a small business switch to the cheapest model for everything?

No. Route by tier: bulk routine work to the cheapest competent model, standard work to the mid-tier, and only brand-critical or high-stakes work to flagship models. Running everything on the flagship is the most common way small businesses overpay for AI.

Do cached input prices really matter?

Yes, especially for agent workloads. Agents re-read the same instructions, templates, and documents repeatedly, and cached input now costs a fraction of fresh input — Terra’s cache reads are $0.05 versus $1.00 per million fresh input tokens. Structuring prompts with stable context first and enabling caching can cut agent bills substantially.