Mostly no. Anyone can chat with Claude Sonnet 5.5 free on Claude.ai, but Claude Code is not on the Free plan, and a chat seat cannot start a run. The API is prepaid, and no official page offers a signup credit. So a Crew of coding agents cannot run on Sonnet 5.5 for $0 unless you qualify for one of a handful of programs. Split the bill into lines and the answer gets sharper: the platform line can be $0, because Sharkly is currently free for organizations of up to 10 people; the model line can’t; and the review line never was.
Anthropic shipped Sonnet 5.5 on September 28 at Sonnet 5’s list price, $2 per million input tokens and $10 per million output, and says it runs 30%+ faster and costs up to 30% less per task. For a team already running several agents, that cuts the other way from “free”: cheaper, faster runs put more finished work in front of the same reviewers. Which agent is on what, what finished, what is waiting on a person: none of that got cheaper this week.
TL;DR
Claude.ai Free gives a person Sonnet 5.5 in a chat window: useful for evaluation, useless to a Runtime. The $0 routes that can start an unattended run are eligibility-gated: Claude Team for scientists, and API credits for researchers and venture-backed startups. Everyone else pays, either Claude Pro from $17 a month with Claude Code switched to Sonnet 5.5, or the API at $2 and $10 per million tokens, where the effort setting moves the bill more than the model does.
Three lines on the bill
A Crew’s cost is three lines that move independently, and “free” only ever applies to one of them at a time.
| Line | What it pays for | Who charges it | Can it be $0? |
|---|---|---|---|
| Seat | The platform where work is assigned, tracked and accepted | Sharkly | Yes, up to 10 people |
| Token | The Runtime’s model usage | Your model provider | Only through the programs below |
| Review | A person reading and accepting the result | Your team’s hours | No |
The seat line. Sharkly’s pricing is $0 for organizations of up to 10 people and $7 per user per month billed yearly above that. Guests, Agents, Crews, Computers and Runtimes do not occupy seats, so adding Agents or Computers never moves this line; hiring the eleventh person does. Plan mechanics: Sharkly pricing explained.
The token line. Sharkly does not sell tokens. Model usage continues through the subscriptions or API keys configured in those tools, which is the point of bring-your-own-key. For Sonnet 5.5 that means $2 in and $10 out per million tokens, with cached reads at $0.20, per Anthropic’s pricing page.
The review line. Agents research, execute, test, and report. People set direction, grant authority, and accept the result. No provider discounts the second sentence, and a faster, cheaper model produces more finished runs per reviewer, so this line grows exactly when the token line shrinks.
What free Sonnet 5.5 access exists, and what it can drive
These are the routes on official pages as of September 29, with the one question that decides whether each is useful to a Crew.
| Route | Who qualifies | Can it start an unattended run? |
|---|---|---|
| Claude.ai Free | Anyone | No. Chat on web, desktop and mobile; Claude Code reads “No” in the Free column; no API key |
| GitHub Copilot Pro, free | Verified teachers and maintainers of popular open-source projects | Only inside Copilot’s own agent and CLI, billed at list price from about $15 a month of AI credits. Not a credential Claude Code can use |
| Claude Team for scientists | PIs at accredited universities and nonprofit research institutes; a limited number of groups, for 12 months | Yes, within that plan’s usage. Standard seats are free and include Claude Code |
| External Researcher Access Program | AI safety and alignment researchers | Yes. $1,000 in API credits for the standard model suite |
| AI for Science | Any researcher can apply, per project | Yes. Up to $50,000 in API credits; biology and chemistry work stays on Opus-class models |
| Claude for Startups | Venture-backed, founded in the last four years, no prior Anthropic credits | Yes, on the first-party API only. The amount is not published |
| Copilot Free and Student, Cursor Hobby, OpenRouter, Vercel AI Gateway’s $5 credit, Google Cloud’s $300 trial, Azure free account | Open to most people | No free Sonnet 5.5 |
The Copilot route is real and metered. Copilot Pro bills Sonnet 5.5 per token at list price, drawing on 1,500 AI credits a month at a cent each: about $15 of tokens by our arithmetic, inside GitHub’s tools, on a gradual rollout. Fine for a maintainer trying the model; it does not authenticate Claude Code on a Computer.
The Team plan for scientists is the only free route that includes Claude Code itself, offered to a limited number of academic and nonprofit research groups for 12 months.
The last row is where free-tier roundups go wrong. Copilot Free and Student reach models only through auto selection, which does not include Sonnet 5.5. Cursor bills it from a pool attached to Pro and above. OpenRouter has no free Claude variant, and Vercel’s $5 credit covers a subset of models that leaves it out. Google’s $300 trial excludes partner models, and Claude in Microsoft Foundry needs a paid Azure subscription. We found no official statement on AWS Free Tier credits for Claude on Bedrock, so we do not count them. For free keys that do exist elsewhere, see free AI API keys.
A chat seat is not a Runtime credential
Four roles, one job each. The Agent defines how work should be handled. The Computer supplies the host and local resources. The Runtime performs the actual agent session, and here that is Claude Code. The Task remains the shared record for the team. An Agent cannot execute anything until a supported Runtime is installed and authenticated on its Computer, and that authentication is a Claude subscription login or an API key configured in the tool.
A Claude.ai Free account is neither. It is a login to a chat window, used by whoever is sitting in front of it. It gives Claude Code no login, because Claude Code is a paid-plan feature, and it comes with no API key, because even a paid chat plan “doesn’t include access to the Claude API or Console,” in Anthropic’s words. So if the question is whether you can run Claude Code agents for free, the chat seat is not the answer. It still has a job.
Use the free chat seat for evaluation with you in the room: paste in a tricky module, read how Sonnet 5.5 plans the change, and decide whether it earns a place in the routing rule at all. One caveat: requests flagged by Anthropic’s cyber or LLM-development classifiers get answered by Sonnet 5 instead, so a security-heavy test may be grading the wrong model.
Use Claude Pro or an API key once work has to begin without anyone watching: a Task picked up from the queue at 2 a.m., a batch of small fixes assigned on Friday afternoon, a teammate’s assignment while you are in review. Those need a credential on a Computer.
The cheapest paid setup for a small Crew
Claude Pro. $17 a month on an annual plan ($200 up front) or $20 billed monthly, and “Claude Code is included in all paid plans.” One trap: Claude Code’s default model on Pro is Opus 5.5, so a Computer running Claude Code is on the flagship unless someone changes it.
# Sonnet 5.5 requires Claude Code v2.1.284 or later
claude --model claude-sonnet-5-5
# or, inside a running session: /model
Pin the full ID rather than the sonnet alias, which per the Claude Code model configuration docs means Sonnet 5.5 only on the Anthropic API and older Sonnets on Bedrock, Google Cloud, Foundry and Claude Platform on AWS. Anthropic describes Pro usage as at least 5x Free’s per five-hour session, with no fixed message count: a subscription has a ceiling, an API key has a bill. The full setup is in how to run Claude Sonnet 5.5 agents in Sharkly.
The API. $2 and $10 per million tokens, prepaid. Credits expire a year after purchase and are non-refundable. In Claude Code, Sonnet 5.5 runs at medium effort by default; the raw API defaults to high.
Effort is the variable that decides this bill. Anthropic’s launch charts put per-task costs next to CursorBench 4.0 scores (Cursor ran the benchmark; Anthropic estimated Sonnet 5.5’s costs at list price):
| Model and effort | CursorBench 4.0 | Cost per task |
|---|---|---|
| Sonnet 5.5, low | 35.8% | $0.50 |
| Sonnet 5.5, medium | 39.2% | $0.70 |
| Sonnet 5.5, high | 47.8% | $1.67 |
| Sonnet 5.5, max | 55.5% | $9.67 |
| Opus 5.5, medium | 52.5% | $2.91 |
By our arithmetic, going from medium to max multiplies Sonnet 5.5’s cost per task by almost 14, while switching to Opus 5.5 at medium multiplies it by about 4. Forty tasks a week come to roughly $28 at medium and $387 at max. The table also puts Opus 5.5 at medium within three points of Sonnet 5.5 at max for under a third of the cost, so the cheaper model at its top setting is not the cheaper choice. Your repository is not CursorBench, but the lesson holds: decide the effort level before you argue about the model. The routing rule is in Sonnet 5.5 or Opus 5.5: which agent gets the task?
Cheap effort also shifts cost onto the review line. Anthropic’s prompting guide says that at low and medium Sonnet 5.5 checks in early on long agentic tasks, and at low it can report code as done without running the tests. In Sharkly a question returns to the Task as a comment, the Task shows waiting for human reply, and it lands in someone’s Inbox. That is a cheap run and an expensive interruption, which is why cost tracking across agents has to count people’s time next to tokens.
Where free ends
We answered the same question for the September 22 models in our breakdown of free and cheap tiers, and Sonnet 5.5 comes out the same shape. Free access is a person in a chat window. Runs cost cents at medium effort and dollars at max. Seats cost nothing up to 10 people. Reading the diff and accepting it has no tier at all.
If you’re already running several agents, Sharkly gives you one place to manage their tasks and results. Start with one Computer whose Claude Code runs Sonnet 5.5 at medium effort, one Agent, and one low-risk Task.
