Blended token prices fell 67% year over year, from $18.40 to $6.07 per million, and 73% of enterprises still overshot their AI budgets (Practical Logix). That pair of numbers is the real answer to "which LLM provider should we choose." The model you pick matters less than how you pick it, because in 2026 all three frontier providers ship capable models within weeks of each other, cut prices against each other, and deprecate products you may have built on. The decision that holds up is the one that assumes you will switch.
We build on all three providers for clients, and we have migrated systems between them. This post compares OpenAI, Anthropic, and Google the way a buyer should: on pricing shape, agent tooling, lock-in, data terms, and fit by workload, as of late August 2026. It does not rank them on benchmark scores, because benchmark leadership changes monthly and rarely predicts performance on your data. It is written for the person signing the build budget, not the engineer choosing a library. Prices quoted are list prices at the time of writing; check the provider pages before you commit, because they will have moved.
Which LLM provider is best for business in 2026?
There is no single best provider; there is a best provider per workload, plus one rule that applies to every business: build behind a model abstraction layer so switching costs days, not months. As a rough guide from our client work, Anthropic's models are our default for agentic and coding workloads where reliability of tool use matters, OpenAI has the broadest ecosystem and the most aggressive pricing this quarter, and Google's Gemini line is the value choice for high-volume, lower-complexity tasks and for companies already standardized on Google Cloud. The rest of this post shows the evidence behind each of those statements and the exceptions to them.
What changed in the last 60 days
- Anthropic released Claude Opus 5 on 24 July 2026 at unchanged Opus pricing of $5 per million input tokens and $25 per million output, with an effort toggle that trades cost against capability per request, plus agent templates and a managed agents offering (Fortune).
- OpenAI cut GPT-5.6 Sol pricing on 24 August 2026 to a promotional $4 per million input and $20 per million output through 21 November (Technology.org). Promotional is the operative word; budget on the list price that returns in December.
- OpenAI announced the sunset of its no-code Agent Builder and hosted Evals for 30 November 2026, directing customers to its Agents SDK (MCP Directory). Anyone who built agents on it has a migration this quarter.
- Google made Gemini 3.7 Flash generally available on 13 August 2026, continuing the Flash line's position as the low-cost, high-throughput option (LLM Gateway timeline). August alone saw 14 model releases across the industry.
- The Model Context Protocol shipped a revised specification on 28 July 2026 with a stateless core and hardened authentication, and all three providers support it, which materially lowers the cost of switching (MCP).
The pattern: a price cut, a deprecation, and two flagship releases in eight weeks. A provider decision made in June is already out of date. Architecture that tolerates that churn is the actual deliverable.
Provider comparison for business buyers
| Dimension | OpenAI | Anthropic | |
|---|---|---|---|
| Frontier list price (input / output per 1M tokens) | GPT-5.6 Sol $4 / $20 promotional to 21 Nov 2026 | Opus 5 $5 / $25; Sonnet tier $3 / $15 | Gemini Pro tier priced competitively; Flash tier well below all frontier models |
| Cost-control features | Model tiers, batch pricing, prompt caching | Per-request effort toggle, prompt caching, batch pricing | Flash tier, context caching, batch pricing |
| Agent tooling | Agents SDK (Agent Builder sunsets 30 Nov 2026) | Agent SDK, managed agents, agent templates | Vertex AI agent tooling, tight Google Workspace integration |
| MCP support | Yes | Yes (originated the protocol) | Yes |
| Platform lock-in risk | Medium; product line changes frequently | Low to medium; API-first, fewer proprietary layers | Medium; strongest when you are already on Google Cloud |
| Enterprise data terms | No training on API data by default; enterprise agreements available | No training on API data by default; enterprise agreements available | No training on paid API data; Vertex AI inherits Google Cloud terms and regions |
| Regional hosting | Expanding; confirm per region | Available via cloud partners; confirm per region | Broadest region coverage through Google Cloud |
| Where we default to it | Broad consumer-facing assistants; teams already on OpenAI | Agents with tools, coding, long documents, regulated workflows | High-volume classification and extraction; Google Cloud shops |
Two honest caveats about that table. First, "no training on API data" is the default for paid API use at all three, but the exact terms, retention windows, and regional options change and differ between API, cloud-marketplace, and enterprise contracts; read the current terms before you sign anything. Second, the pricing column will be wrong within a quarter. That is not a flaw in the table; it is the argument for the architecture in the next section.
The rule that matters more than the provider: build to switch
Every system we ship has three properties that make the provider decision reversible.
- A model abstraction layer. One internal interface for "call a model with these messages and tools," with provider adapters behind it. Switching is a configuration change plus a regression run, not a rewrite. This is a few hundred lines of code and the highest-return engineering hour in the project.
- A provider-neutral evaluation suite. A golden dataset of real inputs and expected outputs, run against every candidate model before and after any change. This is what tells you, with numbers, whether the cheaper model this quarter is good enough for your workload. Without it, provider comparisons are opinions.
- Tools defined once, exposed through MCP. With all three providers supporting the protocol, tool definitions stop being provider-specific. This is why the July 2026 specification revision matters to business buyers, not only to developers.
Companies that skipped these three steps are the ones now migrating off a deprecated builder before 30 November. Companies that did them will re-run their evals against the December list prices and move whatever workloads the numbers say to move. The architecture pattern is described in more depth in building an AI-powered SaaS.
Which provider for which workload?
| Workload | What matters most | Our default in 2026 | Why |
|---|---|---|---|
| Customer support agent with tools (refunds, lookups) | Reliable tool use, refusal behavior, safety under injection | Anthropic Sonnet tier | Strongest tool-use reliability per dollar in our evals; effort toggle for hard cases |
| Document extraction at volume (invoices, forms) | Cost per document, structured output | Google Flash tier, OpenAI small tier as alternate | Lowest cost per million tokens for a bounded task; accuracy gap closes with good prompts |
| Coding agents and internal engineering tools | Long context, multi-step reasoning, tool use | Anthropic Opus or Sonnet | Best sustained agentic coding results in our work; see our tool benchmark below |
| Consumer-facing general assistant | Breadth, ecosystem, familiar behavior | OpenAI | Widest ecosystem; current promotional pricing lowers pilot cost |
| Retrieval over your own documents (RAG) | Long context, citation faithfulness, cost at scale | Anthropic for answer quality, Google Flash for the retrieval-classification steps | Split the pipeline; the expensive model only answers, the cheap one routes |
| Classification, routing, moderation | Throughput and cost | Google Flash or any small tier | Frontier models are wasted here; route with the cheapest model that passes evals |
| Regulated workflows (health, finance, EU data) | Data residency, contractual terms, auditability | Whichever provider offers the region and terms your counsel signs off; often via a cloud marketplace | Terms and regions decide this, not model quality |
Notice how often the answer is "split the pipeline." The single biggest cost mistake we see is routing every request to the frontier model. A well-built system sends 80% to 90% of calls to a small, cheap model and reserves the expensive one for the steps that need it. That design is also what keeps the CIO.com "inference bill nobody budgeted for" from arriving (CIO). Our RAG implementation guide shows the split for retrieval systems specifically, and our benchmark of AI coding tools covers the developer-tooling side, where the provider question shows up as Cursor versus Claude Code versus Copilot.
Pricing: how to compare without being misled
List prices per million tokens are the least useful number in the comparison, for three reasons.
- Output tokens cost four to six times input tokens at every provider, and agentic workloads are output-heavy. A model that is 20% cheaper on input and 10% more verbose on output can cost more per task.
- Caching and batch discounts change the math. Prompt caching cuts repeated-context cost sharply on all three platforms; batch pricing roughly halves cost for anything that can wait. A workload that fits caching on one provider and not another has a different real price.
- Promotions expire. OpenAI's current $4 / $20 rate ends 21 November 2026. Budget on what returns in December, and treat the promotion as a pilot subsidy.
The only reliable comparison is cost per completed task on your own workload, measured in a two-week pilot with the evaluation suite above. That is also the number your ROI model needs; the method is in how to measure AI ROI.
Lock-in: the risk that materialized in 2026
Lock-in used to be a theoretical concern. The Agent Builder sunset made it concrete: teams that built orchestration logic inside a proprietary visual tool now have to rebuild it in code by the end of November. The lesson is not "avoid OpenAI"; every provider deprecates products. The lesson is that orchestration, prompts, tool definitions, and evaluation data must live in code you own, in your repository, with the provider reduced to an API call behind an adapter. When we help clients choose a vendor for an agent build, this is question six on the list in how to choose an AI agent development company, and it is the question most often answered badly.
Security posture also varies with lock-in. The more of your logic sits inside a provider's proprietary layer, the less you can audit and test it against the controls in our AI agent security checklist. Code you own is code you can test.
Frequently asked questions
Is Claude or ChatGPT better for business?
It depends on the workload. In our client work, Anthropic's Claude models are the default for agents that use tools, for coding, and for long-document work, because tool-use reliability per dollar has been strongest in our evaluations. OpenAI's models are the default for broad consumer-facing assistants and for teams already invested in its ecosystem, and its current promotional pricing lowers pilot costs. Run both against your own evaluation set before deciding.
Which LLM is cheapest for business use?
Google's Gemini Flash tier and the small-model tiers from OpenAI and Anthropic are the cheapest per token for bounded tasks such as classification and extraction. For complex agentic work, cost per completed task matters more than cost per token, and that must be measured on your workload. Promotional rates, such as OpenAI's through 21 November 2026, should be excluded from long-term budgets.
Should we use more than one LLM provider?
Yes, at the pipeline level. Most production systems route routine steps to a cheap model and reserve a frontier model for the hard steps, which frequently means two providers. A model abstraction layer and a provider-neutral evaluation suite make this routine rather than risky.
How do we avoid LLM vendor lock-in?
Keep orchestration, prompts, tool definitions, and evaluation data in your own code, put every provider behind a single internal adapter interface, expose tools through the Model Context Protocol, and re-run your evaluation suite against alternative models each quarter. The November 2026 sunset of OpenAI's Agent Builder is the case study in what happens otherwise.
Do the providers train on our data?
By default, none of the three train on paid API data, and all offer enterprise agreements with retention and regional commitments. The terms differ between direct API, cloud-marketplace, and enterprise contracts, and they change. Have counsel read the current terms for the specific channel you buy through, particularly for regulated data.
Key takeaways
- Token prices fell 67% in a year and 73% of enterprises still overshot budgets. The provider matters less than the architecture that lets you switch.
- Late-August 2026 defaults from our client work: Anthropic for agents with tools and coding, OpenAI for broad assistants and ecosystem, Google Flash for high-volume bounded tasks and Google Cloud shops.
- Compare cost per completed task on your workload, not list price per token. Output tokens, caching, and batch discounts change the answer.
- Split pipelines: send 80% to 90% of calls to a cheap model and reserve the frontier model for steps that need it.
- Build to switch: abstraction layer, provider-neutral evals, MCP-defined tools. The Agent Builder sunset is what skipping this costs.
- Data terms and regions, not model quality, decide the provider for regulated workloads. Read the current contract for your purchase channel.
Get a provider recommendation for your workload
If you are choosing a model provider for a build, send us the workload, expected volume, and any data-residency constraints. We will return a one-page recommendation with a pipeline split, an estimated cost per task on current pricing, and the abstraction approach that keeps the decision reversible. We have no reseller relationship with any of the three, so the recommendation is the one we would make for our own AI development projects.
Sources
- Practical Logix: The AI token bill comes due (2026 enterprise FinOps)
- Fortune: Anthropic debuts Claude Opus 5
- Anthropic API pricing
- Technology.org: OpenAI GPT-5.6 Sol price cut for developers
- MCP Directory: OpenAI Agent Builder deprecation
- LLM Gateway: 2026 model release timeline
- Model Context Protocol: July 2026 specification
- CIO: The inference bill nobody budgeted for