The fastest way to reduce AI agent costs is to stop running every step on your most expensive model. Use a frontier model to plan a task and handle the hard judgement calls, then run the routine steps (classifying, extracting, formatting, calling tools in a known order) on a small model that costs a fraction as much. In the worked example below, that one change cuts the bill by close to 90%.
The second way is to make sure you can always switch. Model prices, availability and terms change on the provider’s schedule, not yours. On 7 October 2026 Anthropic released a small model at a tenth of the price of the one before it. A year earlier, it went the other way: Claude Haiku 4.5 launched at 25% more than Claude Haiku 3.5, and when Haiku 3.5 was retired in February 2026, Haiku 4.5 was the recommended replacement. Teams that can move a step to another model in an afternoon keep the savings and avoid the hikes. Teams wired to one provider take whatever comes.
Shogo, the AI agent platform we build, is designed around both ideas. It’s a model-agnostic harness: it connects to models from every major provider instead of locking you to one, sorts every request by how complex it is and sends it to an economy, standard or premium model, and offers its own small model, Hoshi 1.0, at $0.30 per million output tokens.
This is a follow-up to our guide to AI agent pricing models, which covers how platforms charge you. This one covers what you can control: which model does which work, and how free you are to change it.
Why AI agent bills get expensive
A chatbot answers once. An agent works in loops: it plans, calls a tool, reads the result, decides what to do next and repeats, often a dozen times for one task. Every loop sends the model the instructions, the history and the latest result again. If each of those calls goes to a top-tier model, you pay top-tier prices for steps a much smaller model could do just as well.
The price gap between tiers is large. In OpenAI’s own current list, output from its top model costs $50 per million tokens and output from its smallest costs $0.50. That’s a 100x difference from the same company.
What AI models cost in 2026
Prices per million tokens, from each provider’s pricing page on 12 and 14 October 2026 (standard rates, short context):
| Tier | Model | Input | Output |
|---|---|---|---|
| Frontier | gpt-6-astra (OpenAI) | $10.00 | $50.00 |
| Frontier | Claude Fable 5.1 (Anthropic) | $10.00 | $50.00 |
| Frontier | Claude Opus 5.5 (Anthropic) | $4.00 | $20.00 |
| Mid | gpt-6-sol (OpenAI) | $2.00 | $10.00 |
| Mid | Claude Sonnet 5.5 (Anthropic) | $2.00 | $10.00 |
| Small | Claude Haiku 4.5 (Anthropic) | $1.00 | $5.00 |
| Small | Gemini 3.1 Flash-Lite (Google) | $0.25 | $1.50 |
| Small | Claude Haiku 5.5 (Anthropic) | $0.10 | $0.50 |
| Small | gpt-6-luna (OpenAI) | $0.10 | $0.50 |
| Small | Hoshi 1.0 (Shogo) | $0.15 | $0.30 |
Sources: OpenAI, Anthropic, Google, Shogo Hoshi. Claude Haiku 5.5’s price applies to prompts up to 100,000 tokens; longer prompts cost $0.50 input and $2.50 output.
Per-token prices aren’t perfectly comparable across providers, because each model splits text into tokens differently. Anthropic notes that its Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text.” Test your own prompts before you compare bills.
Newer isn’t more expensive, and older isn’t cheaper
The saving comes from using a smaller tier, not an older model. Claude Opus 4.1 cost $15 input and $75 output per million tokens until Anthropic retired it on 5 August 2026. Claude Opus 5.5, released since, costs $4 and $20. Teams that stayed pinned to the model they started with paid almost four times more for the older model, right up until it was switched off.
Plan with a frontier model, run the steps with a small one
Most agent tasks split cleanly into two kinds of work.
Planning and judgement is where a frontier model earns its price: working out what a vague request actually needs, breaking a job into steps, handling an edge case nobody wrote a rule for, and checking the final result.
Execution is everything after the plan exists: pulling fields out of an invoice, tagging a ticket, reformatting data, translating, and calling the next tool in a known sequence. Google describes its Flash-Lite model as built for exactly this: “high-volume agentic tasks, translation, and simple data processing.”
| Use a frontier model for | Use a mid-tier model for | Use a small model for |
|---|---|---|
| Planning a multi-step task | Drafting a customer reply | Classifying and tagging |
| Ambiguous or unusual requests | Summarising a long document with nuance | Extracting fields from documents |
| Changes across many files of code | Writing a single function | Formatting and cleaning data |
| Legal, financial or medical judgement | Answering from a knowledge base | Translation |
| Final review of the agent’s work | Tool calls in a plan that already exists |
A useful rule: if you could write the step as a clear instruction for a new hire on their first day, a small model can probably do it. If you’d want your most senior person to look at it, keep it on a frontier model.
The worked example: the same agent at three costs
Take an agent that handles 1,000 tasks a day. Each task reads about 10,000 tokens (instructions, history, documents) and writes about 1,000. Here’s what 1,000 tasks cost on each model, using the prices above:
| Setup | Cost per 1,000 tasks | Cost per month (30,000 tasks) |
|---|---|---|
| Everything on gpt-6-astra | $150.00 | $4,500 |
| Everything on Claude Opus 5.5 | $60.00 | $1,800 |
| Everything on Claude Haiku 4.5 | $15.00 | $450 |
| Everything on Hoshi 1.0 | $1.80 | $54 |
| Everything on gpt-6-luna | $1.50 | $45 |
| Everything on Claude Haiku 5.5 | $1.50 | $45 |
| 10% on gpt-6-astra, 90% on gpt-6-luna | $16.35 | $490.50 |
| 10% on gpt-6-astra, 90% on Hoshi 1.0 | $16.62 | $498.60 |
| 10% on Claude Opus 5.5, 90% on Hoshi 1.0 | $7.62 | $228.60 |
The mixed rows assume planning and review take 10% of the work and the rest is routine. On that assumption, keeping the frontier model only for planning cuts the bill by 87% to 89% compared with running everything on it. Your split will differ, so measure it: log which steps use which model for a week before you change anything.
Running everything on a small model is cheapest of all, but you give up the planning quality that keeps an agent from going wrong on unusual requests. The mixed setup is usually the better trade.
Research backs this up
Sending each request to the cheapest model that can handle it is called model routing (or LLM routing), and it has been studied for several years.
- FrugalGPT (Chen, Zaharia and Zou, 2023) tested sending each query through a chain of cheaper models first and found it could “match the performance of the best individual LLM (e.g. GPT-4) with up to 98% cost reduction.” (paper)
- RouteLLM (Ong et al., 2024) trained routers that choose between a stronger and a weaker model per request and reduced costs “by over 2 times in certain cases” without lowering response quality. (paper)
Both studies used 2023 and 2024 models, so treat them as proof the approach works rather than as a forecast of your savings.
Three more ways to lower your AI agent bill
Model routing is the biggest lever. These three stack on top of it:
- Cache what repeats. Agents send the same instructions on every call. Providers charge far less for cached input: Claude Haiku 4.5 cache hits cost $0.10 per million tokens against $1.00 for fresh input, and Hoshi 1.0 cached input is $0.01.
- Batch anything that can wait. Anthropic’s Batch API gives “a 50% discount on both input and output tokens,” and OpenAI’s batch prices are half its standard ones. Overnight reports and bulk tagging don’t need instant answers.
- Send less context. Most of an agent’s input is instructions and history it doesn’t need for the current step. Our guide to tokenmaxxing covers trimming system prompts and loading skills only when a task needs them.
Avoid vendor lock-in: why a model-agnostic harness protects your budget
Matching the model to the task only saves money if you’re free to pick the model. That freedom is the job of the harness.
What is an AI agent harness?
An AI agent harness is the layer around the model: it holds your agent’s instructions, tools, memory, permissions and routing rules, and sends each step to whichever model you choose. In a model-agnostic harness, the model is a part you can swap. In a single-provider tool, the model is the foundation, and changing it means rebuilding.
The difference only shows when something changes at the provider. In the last two years, plenty did.
What changed at the major AI providers in 2025 and 2026
Every item below comes from the provider’s own documentation or a named news report:
| What happened | Example | What it means for a single-provider agent |
|---|---|---|
| Models retired on a fixed schedule | Anthropic deprecated Claude Opus 4.1 on 5 June 2026 and retired it on 5 August 2026: 61 days. On 14 October 2026, OpenAI announced that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will be removed from its API on 14 January 2027. | Every agent built on those models has to be migrated and re-tested by the deadline. Anthropic promises “at least 60 days’ notice” for public models; OpenAI promises at least 6 months for generally available models and as little as 2 weeks for preview models. |
| The replacement costs more | Claude Haiku 3.5 ($0.80/$4) was retired on 19 February 2026, and Anthropic’s recommended replacement, Haiku 4.5, costs $1/$5: 25% more. When Google made Gemini 2.5 Flash generally available in June 2025, input went from $0.15 to $0.30 per million tokens, and the preview it replaced was switched off on 15 July 2025. | A forced move can be a forced price rise. |
| Access cut with days of warning | In June 2025, Windsurf said Anthropic gave it “less than five days of notice” before cutting “nearly all of our first-party capacity to all Claude 3.x models,” as reports said OpenAI was buying Windsurf. | Business decisions you can’t see coming can remove a model from your product. Windsurf had to move to third-party capacity at short notice. |
| The product you built on is shut down | OpenAI’s Assistants API was removed on 26 August 2026. Its Agent Builder was deprecated on 3 June 2026 and is scheduled to shut down on 30 November 2026. | Agents built inside one vendor’s builder have to be rebuilt elsewhere, not just pointed at a new model. |
| Outages | On 11 December 2024, OpenAI’s status page reported that between 3:16 PM and 7:38 PM PST “all OpenAI services experienced significant degradation or complete unavailability.” | For four hours and 22 minutes, any agent that could only call OpenAI had no other model to move to. |
None of the major labs has gone out of business. But for an agent that depends on one of them, a retired model, a withdrawn product or a cut-off contract has the same effect as a closure: the agent stops until someone rebuilds it.
Lock-in costs you twice: when prices rise, and when they fall
Lock-in isn’t only a risk when things go wrong. Prices also fall, and only teams that can switch collect the saving. Claude Haiku 5.5, released on 7 October 2026, costs $0.10 input and $0.50 output per million tokens, a tenth of Claude Haiku 4.5’s price. Anthropic describes it as built for “summarization, classification, routing” in agent systems, with a larger model planning the work and handing sub-tasks down. A team that can point its routine steps at a new model the week it ships cuts that part of its bill by 90% straight away. A team that can’t keeps paying the old rate.
Most enterprises already use several models
In a16z’s 2025 survey of 100 CIOs across 15 industries, 37% of respondents were using five or more AI models, up from 29% a year earlier. The main reason wasn’t fear of lock-in but that models are better at different jobs: “model differentiation by use case has become increasingly pronounced and is the main reason enterprises buy models from multiple vendors.”
The same survey found that switching is getting harder as agents get more complex. One leader told a16z: “All the prompts have been tuned for OpenAI… changing models is now a task that can take a lot of engineering time.” That’s the problem a harness solves: when instructions, tools and memory live in a layer you control rather than inside one provider’s product, switching becomes a test-and-compare job, not a rebuild.
Six questions to ask any AI agent platform
- Does it work with models from more than one provider, including open-weight or local models?
- Can different steps of the same agent use different models?
- How quickly are new models added after release?
- If a model you use is retired, what do you have to change: a setting, or your agent?
- Does it send simple requests to a cheap model and hard ones to a strong model automatically, and can you choose which models it uses?
- If you leave, can you take your agents’ instructions, tools and data with you?
How Shogo cuts AI agent costs without locking you in
- Auto-route picks the right tier for every request. With auto-route on (“Shogo routes every task to the best model for it,” as the homepage puts it), a fast rule-based classifier reads each request and sends it to one of three tiers: economy, standard or premium. A short lookup like “check whether the README has install instructions” goes to economy; “debug why login fails” goes to premium. It decides in under a millisecond without an extra model call, so the routing itself adds no cost.
- One key, every major provider. Shogo routes across Anthropic, OpenAI, Google, xAI and local models through one Shogo API key, so you’re never tied to a single provider. The homepage puts it as “the best from every lab, the day it ships.”
- The model is a setting, not a rebuild. Your agents’ instructions, tools and memory live in Shogo, not inside one provider’s product. You can “swap in a click, or let Shogo pick the best one for each task,” so a retired model, a price rise or a cheaper new release means changing a model choice, not rebuilding the agent.
- Built-in safeguards keep cheap routing safe. When the classifier isn’t confident a request is simple, it rounds up to the standard tier. Long conversations and precision-sensitive work, such as deployments and database operations, always get at least the standard tier. If a sub-agent’s task fails or comes back empty, Shogo retries it on a higher tier.
- You decide which model sits in each tier. Admins can assign any available model to the economy, standard and premium tiers, so when a cheaper model ships, like Claude Haiku 5.5 this month, you can move a whole tier to it without touching a single agent.
- No lock-in to us, either. Hoshi 1.0 uses the standard OpenAI request format. The Hoshi page says you can “switch back or route around us anytime,” with “nothing proprietary to unwind.” Shogo usually runs on Shogo’s cloud, and can be deployed in your own cloud, a VPC or on-premises, alongside local models.
- Hoshi 1.0 for routine work. Shogo’s own model costs $0.15 per million input tokens, $0.01 cached and $0.30 output. That low output and cache price suits agents, which re-read the same context on every loop: with 80% of input cached, 1,000 tasks from the worked example cost $0.68 on Hoshi against $0.78 on gpt-6-luna or Claude Haiku 5.5. In Shogo’s published benchmarks it scored higher than Claude Haiku 4.5 and gpt-5.4-nano on all three evals, and it solved all 83 tasks of the Aider Polyglot coding benchmark for $0.41 in total, about half a cent per task. It’s also available directly through an OpenAI-compatible API.
- Flat pricing on top. Paid plans are unlimited within fair-use windows rather than a credit pool, so cheaper routing stretches your plan instead of shrinking a bill. Pro is $20 per seat per month and includes all models.
FAQ
How can I reduce my AI agent costs?
Run planning and judgement on a frontier model and routine steps on a small one. In our worked example, keeping a frontier model for 10% of the work and a small model for the rest cut the bill by 87% to 89%. Caching repeated instructions, batching work that can wait and trimming unnecessary context lower it further.
Do AI agents need a frontier model for every task?
No. Frontier models are worth their price for planning multi-step tasks, handling unusual requests and reviewing results. Classifying, extracting, formatting, translating and running tool calls in an existing plan can run on small models that cost a fraction as much.
Which AI agent tasks can run on cheaper models?
Tasks you could write as a clear instruction: tagging tickets, extracting invoice fields, reformatting data, translation, summarising short documents and calling the next tool in a known sequence. Keep ambiguous requests and high-stakes judgement on a frontier model.
What is AI model routing?
Model routing, or LLM routing, sends each request to the cheapest model that can handle it instead of sending everything to one model. Research such as FrugalGPT (2023) found it can match the best single model’s performance at up to 98% lower cost on the tasks tested.
Which AI agent platforms can use multiple LLM models?
Look for platforms that work with models from more than one provider and let you use different models for different steps. Shogo routes across Anthropic, OpenAI, Google, xAI and local models through one API key, adds new models as they ship, and with auto-route on sends each request to an economy, standard or premium model based on how complex it is. Framework tools you host and code yourself can also call several providers, but you set up the routing.
What is an AI agent harness?
An AI agent harness is the layer around the model that holds an agent’s instructions, tools, memory, permissions and routing, and sends each step to a model. In a model-agnostic harness such as Shogo, the model is a replaceable part: you can change it, or use several at once, without rebuilding the agent.
How do I avoid vendor lock-in with AI models?
Keep your agents’ instructions, tools and data in a layer you control rather than inside one provider’s product, use a platform that works with several providers, and keep a tested second-choice model for each important step. Then a retirement, price change or outage at one provider means changing a setting rather than rebuilding.
What happens to my AI agent when its model is retired?
Requests to a retired model fail, so the agent stops working until it’s moved to another model. Anthropic gives at least 60 days’ notice for public models and OpenAI at least 6 months for generally available ones (as little as 2 weeks for preview models). Anthropic retired Claude Opus 4.1 on 5 August 2026, 61 days after announcing it.
Is a model-agnostic AI agent platform worth it?
For most teams, yes. Being able to switch lets you use a cheaper model the week it’s released, avoid a price rise on a replacement and move to another provider when one has a problem. In a16z’s 2025 survey of 100 CIOs, 37% of respondents already used five or more models.
Is an older AI model always cheaper?
No. Cheaper models are smaller tiers, not older versions. Claude Opus 4.1 cost $15 input and $75 output per million tokens until it was retired in August 2026, while the newer Claude Opus 5.5 costs $4 and $20.
What is Hoshi 1.0 and how much does it cost?
Hoshi 1.0 is Shogo’s own AI model for agent work. Through its API it costs $0.15 per million input tokens, $0.01 per million cached input tokens and $0.30 per million output tokens, and it works with OpenAI’s SDKs by changing the base URL, key and model name. In Shogo’s benchmarks it scored higher than Claude Haiku 4.5 and gpt-5.4-nano.
What does Shogo cost?
Free to start, then Basic at $8 per month, Pro at $20 per seat per month and Business at $40 per seat per month. Paid plans are unlimited within fair-use windows, with no credit pool. Pro and above include all AI models.
Sources
- OpenAI, API pricing, checked 12 October 2026.
- Anthropic, Claude pricing, checked 12 and 14 October 2026.
- Google, Gemini Developer API pricing, checked 12 October 2026.
- OpenAI, Deprecations, checked 14 October 2026.
- Anthropic, Model deprecations, checked 14 October 2026.
- Anthropic, Claude Haiku (Claude Haiku 5.5, 7 October 2026).
- Google, Gemini 2.5: Updates to our family of thinking models, 17 June 2025.
- TechCrunch, Windsurf says Anthropic is limiting its direct access to Claude AI models, 3 June 2025.
- OpenAI status page, API, ChatGPT & Sora Facing Issues, 11 December 2024.
- a16z, How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025, survey as of 8 May 2025.
- Chen, Zaharia and Zou, FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance, 2023.
- Ong et al., RouteLLM: Learning to Route LLMs with Preference Data, 2024.
- Shogo, Hoshi 1.0 API, Shogo Launches Hoshi 1.0 and How We Get to 12x.
Related reading: AI agent pricing models explained, tokenmaxxing and your AI budget, why Shogo and Shogo pricing.
Prices checked against each provider’s pricing page on 12 and 14 October 2026; retirement dates checked against each provider’s deprecations page on 14 October 2026. Model prices and schedules change often; check the linked pages before budgeting.
Last reviewed and updated: October 2026.
