Freelance AI engineers in 2026: rates, platforms and vetting

A freelance AI engineer costs $35 to $60 an hour on Upwork and $60 to $100+ an hour through Arc, by each company’s own figures as of 30 October 2026. In UK contract adverts for AI engineers, the median day rate as of 30 October 2026 was £600, about $800. You can hire one on an open marketplace such as Upwork or Contra, through a vetted network such as Toptal, Arc, Lemon.io or Braintrust, or from community threads on Hacker News. A freelancer fits a bounded task when someone on your side can define a good answer and owns the architecture.

Both sides of the market have grown in a year. IT Jobs Watch counted 550 UK contract adverts for AI engineers in the six months to 29 October 2026, up from 204 a year earlier, and 61% of them mentioned large language models (LLMs).¹ In Hacker News’s monthly “Who wants to be hired?” thread, posts that offer contract or freelance work and list LLM skills went from 22 in October 2025 to 151 in October 2026, by a keyword count of the public threads. Only 36 of the 151 mention evals, the test sets that show whether an AI feature still works after a change.

9io wrote this post. We’re a product studio that also supplies senior AI engineers by the hour, so we’re an alternative to a freelancer, and the decision rule near the end says when a freelancer is the better choice. Figures come from public data and each company’s own pages, read on 30 October 2026, and claims such as “top 2%” are the companies’ own. Our guide to hiring AI engineers covers employing one, and the agency and managed-team routes.

Freelance AI engineer rates in 2026, from public sources

Published rates measure different things. Some describe what clients paid on past contracts, some are a platform’s asking price, and some are rates quoted in job adverts or on candidates’ profiles. The table lists each figure with what it measures, as of 30 October 2026.

The figures sit far apart because they cover different markets. Upwork’s covers every level and country on its marketplace, Arc’s and Lemon.io’s are asking prices for engineers they have screened, and Go Fractional’s covers lead roles taken part-time.

Two details change a budget. As of 30 October 2026, Go Fractional’s benchmark shows employers posting $105 an hour on average for this role, while candidates asked $166. A budget set from job posts can miss the people whose profiles set the higher figure, although 10 job posts is a small sample. On marketplaces, fees of up to 15% also come on top of the freelancer’s rate, as the platform table below shows.

UK rates are rising. As of 30 October 2026, IT Jobs Watch’s median for AI engineer contracts was 14% higher than a year earlier, and the median for all UK IT contract adverts was £525. The 10th to 90th percentiles for AI engineers ran from £418 to £875 a day, about $560 to $1,170 at the Federal Reserve’s rate of $1.3375 to the pound on 23 October 2026.⁸ Over an eight-hour day, £600 is about £75, or $100, an hour.

Freelance AI engineer salary: what a year of billing adds up to

A freelancer has no salary. Gross income is the hourly rate times the hours billed, and platform fees, tax, insurance, equipment and the unpaid time spent finding the next client all come out of it. We found no public source for the billable hours of freelance AI engineers, so the table shows the arithmetic for two assumed totals, at rates chosen to span the ranges above.

A full-time year is about 2,080 paid hours, so 1,500 billed hours leaves about 580 for holidays, sales and admin. On Upwork, the freelancer’s own service fee of 0% to 15% per contract, as of 30 October 2026, comes out of these totals.⁹

For comparison, Robert Half’s 2027 Salary Guide puts the starting salary for a US AI engineer at $154,500 to $196,750 as of 30 October 2026, before the employer adds payroll taxes and benefits.⁷ AI skills also carry a premium. In Robert Half’s survey for the guide, 92% of technology leaders said they were offering higher pay for relevant AI skills.¹⁰

Where to find freelance AI engineers, and what each platform charges

The table compares the main places to hire freelance AI engineers, from each company’s own pages as of 30 October 2026. Vetting figures are the companies’ own claims, and “not published” means we couldn’t find the figure on their public pages.

On $10,000 of freelancer billing, the client’s fee as of 30 October 2026 is $500 on Upwork Basic plus the one-time contract fee, $1,000 on Business Plus and $1,500 on Braintrust. Contra charges $29 on a payment of that size, plus about $290 of processing by card or up to $5 by an approved bank transfer. Arc and Lemon.io quote one hourly price, and Toptal adds its subscription to a rate it quotes, so compare their all-in hourly prices.

Upwork says it has 18 million-plus freelancers in 180 countries and verifies each one’s identity and location.⁹ Its AI engineers carry a 4.8 out of 5 rating across 12,618 client reviews, and the screening is yours to do. Upwork’s own hiring guide tells clients to define how model performance will be measured and accepted before the project ends, and to agree who owns the datasets, trained models, prompts and source code.³ Our post on Upwork alternatives prices the hours that screening takes.

Contra describes hiring on its platform as “100% commission-free”, yet as of 30 October 2026 its client help centre lists the flat platform fee shown in the table, waived when the freelancer pays for Pro.¹⁴¹³ Its pages are organised around design and no-code tools, although it has category pages for AI developers.

Braintrust’s fee terms, last updated on 19 July 2023, add 15% to what the client pays, so as of 30 October 2026 a $90 hourly rate costs $103.50.¹⁶ Its marketplace page promises “top-2% vetted talent in under 48 hours” from 2 million-plus members, and its pricing page says talent pay no platform fee.¹⁸¹⁷

Toptal, Arc and Lemon.io screen engineers before you meet them. Toptal says an AI developer hire typically takes about 48 hours, Arc says it matches a freelance AI developer in about 72 hours, and Lemon.io offers two or three matched candidates within 24 to 48 hours.²¹⁴⁵ Our Toptal alternatives post compares these networks in more detail.

Turing’s talent pages are written for enterprise AI programmes, and its hiring page promises engineer profiles in as little as 72 hours.²⁶ Our Turing alternatives post explains how Turing’s business has changed.

Communities where AI engineers look for contract work

Hacker News runs a “Who wants to be hired?” thread each month, where people post their location, remote preference, technologies and contact details, and the rules say agencies and recruiters are off topic.²⁷ The October 2026 thread had 691 posts by 30 October. A keyword count found 282 that mention freelance, contract, consulting, part-time or fractional work, and 151 of those also mention LLM work such as RAG, agentic systems, MCP, evals or fine-tuning. A year earlier the thread had 352 posts, and 22 matched both.²⁸

Only 36 of the 151 mention evals or evaluation, so ask the rest how they test their work. The posts are free to read, and the people in them have asked to be contacted about work. You do all the screening and write your own contract.

The separate “Freelancer? Seeking freelancer?” thread is smaller. The whoishiring account, which posts the site’s official hiring threads, last posted it in October 2025, when it drew 118 comments, and since November 2025 other users have posted it each month. The October 2026 version had 22 comments by 30 October.²⁹

Open-source work is the other public record. Someone who maintains an eval harness or an MCP server has code you can read before the first call, and their issue threads show how they respond when something breaks.

How to vet a freelance AI engineer for production LLM work

Rates and badges don’t show whether someone can keep an AI feature working after launch. The checklist asks for things you can inspect, and it applies whatever the title says, whether you need an LLM engineer or want to hire generative AI engineers for a product feature. Our hiring guide has interview questions on the same skills.

• An eval set they built. Where the cases came from, how many there are, the pass mark, and a regression it caught. Ask them to run it against a second model and explain the difference.
• Retrieval measured on its own. Labelled queries with the documents that should come back, recall measured before any prompt tuning, and the user’s permissions applied before ranking. They should also know when long context or a database query is enough, which our RAG vs fine-tuning vs long context comparison covers.
• Agents with limits. Which tools an agent can call, which actions wait for a person’s approval, and what happens when a retrieved document contains instructions. OWASP’s Top 10 for LLM applications (2025) puts prompt injection first, and also lists excessive agency and unbounded consumption.³⁰
• A system still running in production. Its cost per request, its 95th-percentile latency, what alerts someone when it fails, and the last thing that broke.
• Spend limits. How they cap spend per job and per day. Provider rate limits don’t stop a job that runs steadily below them.
• A plan for model retirements. How they would move to a replacement and show that nothing got worse. Anthropic gives at least 60 days’ notice, and as of 30 October 2026 Claude Sonnet 4.5 retires on 30 November.³¹ OpenAI shuts down gpt-4o, gpt-4o-mini, gpt-4.1 and gpt-4.1-mini on 14 January 2027.³² Our one-week routine for model upgrades shows the work involved.
• A written handover. Prompts, eval sets and runbooks committed to your repository, with the reason for each threshold.

Then pay for a short trial task. Pick 30 to 50 real inputs from your product with answers a domain expert would accept, set a budget of a few days, and give the freelancer a model key with a spend cap from your own account. Ask for the eval harness before the feature. Judge the result on the score, the failures grouped by cause and the cost per input. Toptal’s and Arc’s trials of up to two weeks are long enough for a task like this.²⁰²²

What the contract needs: IP, data, API keys and model costs

Freelance contracts are often short templates, so check that these points are in yours. This is general information, and a lawyer should check the contract for your jurisdiction.

Assign the IP in writing. Paying for the work doesn’t by itself make it yours. GOV.UK says someone working under a contract for services “will usually retain copyright” in what they produce unless a contract says otherwise.³³ In the US, a commissioned work counts as a “work made for hire” only if it falls into one of nine categories and both sides sign an agreement saying so, and computer programs aren’t on the list.³⁴ Arc’s terms call freelance work product works made for hire, and then assign it to the client as well in case a court disagrees.²³

The platforms’ defaults differ, as the table shows. Upwork’s default terms pass ownership once the freelancer is paid in full, and make the freelancer list background technology and third-party material with each delivery.¹¹ Contra’s terms say it doesn’t transfer ownership of any work product, so your own agreement decides.¹⁵ In a direct contract, assign the code, prompts, eval sets, datasets and any fine-tuned model weights, and ask for the same list of third-party material.

Keep data access narrow. A freelancer who handles personal data on your behalf is likely to be your processor under the UK GDPR. The ICO’s guidance says the contract must then include, among other things, that they act only on your documented instructions and delete or return the data at the end.³⁵ Where you can, give them a copy with personal data removed, through accounts you own, and remove access on their last day. Say which AI coding assistants may read your code and data, and under whose account.

Own the model accounts and the bill. Keep provider accounts in your company’s name and give each freelancer their own key. OpenAI’s help centre says to use a unique key for each team member, says that sharing keys is against its terms of use, and recommends keys with an expiry date.³⁶ It also recommends monthly spend limits with hard enforcement, because spend alerts alone don’t stop traffic. Anthropic’s workspaces scope keys to one workspace with its own monthly spend limit, and archiving a workspace archives its keys within seconds.³⁷

Pay the provider directly, so the usage history stays with you when the contract ends. Small per-call costs add up without caps. On a fashion commerce platform we build and run, a background job made 419,057 vision calls it didn’t need, costing about $1,100, as our write-up of the bug explains.

Say who reruns the evals. Models change after the contract ends. Name who reruns the eval set when a model is retired, and who pays for that time. OpenAI’s deprecations page, read on 30 October 2026, says its hosted Evals platform becomes read-only on 31 October and shuts down on 30 November 2026, so keep the eval set and the code that runs it in your repository.³²

Check employment status on long engagements. A long, full-time engagement under your direction can make a contractor an employee in law. In the UK, medium and large private-sector clients decide the off-payroll (IR35) status of contractors who work through their own company.³⁸ In the US, the IRS looks at behavioural control, financial control and the type of relationship.³⁹ Arc’s and Lemon.io’s terms say the platform doesn’t supervise the work, so directing the freelancer falls to you.²³²⁵

Freelancer, agency, studio or in-house: a decision rule

Decide who will own the architecture and the evals first, then pick the route.

A freelancer is the wrong choice when nobody on your side can tell a good answer from a bad one. The acceptance test then becomes a demo, and the evals, monitoring and model bill fall outside the contract. With a technical lead in place, the vetted networks above are built for adding one engineer to a team you direct.

A studio is one way to cover the third row. 9io is a small team of senior engineers with a fractional CTO, a designer, a product manager and QA, and you can hire AI engineers from us by the hour. Each proposal names the people, their hours and each person’s rate, work usually starts within a couple of days of agreeing it, and the code lives in your repository from the first commit. As of 30 October 2026, plans start at $5,000. Our post on what a fractional CTO costs has market rates for the last row, and our breakdown of the cost to build an AI agent shows how a project quote is built.

How these figures were gathered

• Company figures come from each company’s pricing, FAQ, hiring and terms pages, read on 30 October 2026. Vetting rates and match times are their own claims, which we haven’t tested, and we didn’t hire through any platform for this post.
• UK rates are the rates quoted in contract adverts that IT Jobs Watch collected over the six months to 29 October 2026. Agreed rates can differ from advertised ones.
• Hacker News counts use each thread’s top-level posts from the public Hacker News API on 30 October 2026.⁴⁰ A post counted as contract work if it mentioned freelance, contract, consulting, part-time or fractional work, after phrases such as “API contracts” and “smart contracts” were removed. It counted for LLM skills if it named LLMs, RAG, agentic systems, LangChain, LangGraph, LlamaIndex, MCP, evals, fine-tuning or embeddings. Keyword matches are approximate.
• Currency conversions use the Federal Reserve’s H.10 rate of $1.3375 to the pound on 23 October 2026.
• The income table uses assumed billed hours. The post leaves out rates in other countries beyond the platforms’ worldwide figures, quality differences between individuals, and agency or managed-team rates, which our hiring guide compares.

Frequently asked questions