Claude Code for GTM: How It Works, What It Costs, What It Got Wrong
What Claude Code does in a month of GTM research, drafting and checks, what the plans cost as of November 2026, and the dated failures that show where a person has to say yes.
One 37-agent run on 4 July shows what happens with Claude Code for GTM when nobody picks the model. Every agent inherited the session model, the top-tier one. The run used about 830,000 tokens in 4 minutes and hit the session usage limit before most of the agents had run. A re-run on lighter tiers finished.
That run shaped how we use Claude Code. It lives in your terminal, reads files, runs commands and calls tools, and we point it at account research, first drafts and checks. Our call is to let it read and draft first, to have a different agent verify the work, and to keep a person on anything that reaches a remote system, whether that's a send, a purchase, a cancellation or a publish.
How Claude Code works
Claude Code runs a loop with three phases that blend together: gather context, take an action, check the result. It chains dozens of these steps and corrects course from what each one returns. You can interrupt it at any point. All of this comes from Anthropic's documentation, as of November 2026.
Built-in tools come in five groups: file operations, search, shell commands, the web and code intelligence. It can also spawn subagents. Each works in its own context window and returns a summary. On top sit skills, written procedures that load in full only when used, plus MCP servers and hooks. The model reasons; tools do the acting, and the permission mode sets which actions need your yes.
Undo has a boundary. Claude Code snapshots a file before editing it, so file edits can be rewound. But actions that touch remote systems, such as databases, APIs and deployments, can't be checkpointed. The permission mode is your control there.
Anthropic lists four modes: Manual, Accept edits, Plan and Auto. Manual asks before file edits and shell commands, while Plan explores and proposes without editing. In Auto, a classifier reviews most actions in the background and blocks risky ones. And context compacts into a summary as a session fills, so detailed instructions from early on can be lost. Rules that must persist belong in a CLAUDE.md file.
What Claude Code for GTM looks like in a normal month
| Job | One dated example | Where a person sits |
|---|---|---|
| Account research | 3 November: one prospect dive, our name for a full research file on one account. A 34-page PDF from six research lanes, each an agent with a written brief, of about 15 minutes each, with discovery on built-in web search (about 85 main calls and 26 inside the lanes). | A separate verifier confirmed it and we applied 8 wording fixes. A person decides who reads it and whether anything is sent. |
| Employment checks | Since 26 September, only a contact confirmed as still at the company is sendable. | A lookup that can't settle it returns UNVERIFIED, and a human opens the profile. |
| Drafting articles | 27 October: 12 drafts, each run through draft, verify, fix, verify, fix and a fresh recheck by a third agent. All 12 ended READY. | All 12 synced to the site as drafts, six of them in a later pass after the first sync step stopped (see the log below). Nothing was published. A person approves each publish. |
| The verify loop | 22 October, one article. Round one found 1 blocking, 9 structural and 16 wording problems. Round two found 0 blocking, 7 structural and 9 wording, and still said NOT READY. A fresh recheck said READY. | The verifiers are other agents. A person reads the article before it goes live. |
| A tool server | 31 October: reads across a whole cold-email infrastructure account, then 4 mailbox renames, a profile picture set on every active mailbox in the account, 6 retired domains removed and 5 test sends to inboxes we own. Nothing was bought and nothing went to a prospect. | Each write waited for a person's yes. A separate piece covers the session. |
| Data lookups | 27 October: 591 records from a data vendor on a 200-domain lookup test. | A file lists which vendor serves each function. |
| Reusable skills | 21 skill folders on 6 November, two of them retired stubs. | Plain files, so a person can read what the agent was told. |
The employment rule is written up in employment checks before send. Claygent versus Claude Code for GTM research covers which tool fits which lookup. Wiring Clay's API into a Claude Code agent covers the connection between the two.
What Claude Code costs
Claude Code is included in every paid Claude plan, and it shares the plan's usage limits. Terminal work and chat draw from one pool. These are Anthropic's prices as of November 2026, before tax.
| Plan | Price a month | Usage | Claude Code |
|---|---|---|---|
| Free | $0 | Limited, every 5 hours and every week | No |
| Pro | $17 billed annually ($200 up front), $20 monthly | Baseline | Yes |
| Max | $100 | 5x Pro, every 5 hours and every week | Yes |
| Ultra | $200 | 10x Pro every week, no 5-hour limit for verified accounts | Yes |
| Team Standard seat | $20 a seat billed annually, $25 monthly | More than Pro | Yes |
| Team Premium seat | $100 a seat billed annually, $125 monthly | 5x a Standard seat | Yes |
| Enterprise | $20 a seat billed annually, 20 seats minimum, plus usage at API rates | Scales with model and task | Yes |
At the $20 monthly Pro price, Max and Ultra cost the same per unit of weekly usage: $100 over 5 and $200 over 10 both come to $20. The multiples are Anthropic's, not token counts. Ultra's extras over Max are the end of the 5-hour window for verified accounts and $200 a month in API credit against Max's $100.
At a limit you wait for the window to reset or keep going on usage credits billed at API rates. At the list rates on the same page, the top model, Fable 5.2, costs $8 per million input tokens and $40 per million output. Sonnet 5.5 costs $2 and $10, a quarter of each at list price. We pin a model per stage because the 4 July run ran out of session before it ran out of agents.
The costs that stopped our work were limits. The 4 July run is one. In early June a long audit resume died at a monthly spend cap after about 2.36 million subagent tokens, with six specialists already failed. And in mid-July a 165-agent run hit the session limit mid-run and resumed from its saved results.
The data side is easier to read than the model side. The 22 and 27 October run reports count vendor records, not Claude tokens. The report for the 27 October run even says the model used per lane was not passed to it.
What a dive or an article costs us in dollars, we haven't measured. We haven't priced our own seat against it either, or read our Anthropic invoice, so nothing here says what we pay. The test still to run is small: read the usage fields on one dive and one article, total the tokens, and print the plan price beside them.
What it got wrong, with dates
| Date | What happened | What we changed |
|---|---|---|
| 24 July | Drafting agents' own grounding notes looked confident. An independent verifier with a live fetch found 5 fabrications across 3 drafts, among them an invented pricing-page story, a made-up statistic and a spliced quote. | Independent verify before anything syncs. A statistic counts as confirmed only after a primary-source fetch. |
| 2 September | A second-look check on a diagram returned CLEAN and wrote that all five labels sat inside their circle. Three crossed the outline. It had run on a mid-tier model at low effort and gave a prose read with no numbers. | Second looks check containment by geometry and state the numbers. Never at low effort. |
| 2 and 17 September | Verify agents on long pieces ended without returning output, four times on the 2nd. On the 17th, 3 of 6 fix agents died the same way, and an empty return drops silently as failed. | Fall back to direct agent calls. Recover from disk and the run journal. Never re-run the whole lane. |
| 15 October | A 30-page prospect dive passed the gate and the scan. The independent verifier refused to call it ready: 2 stale facts and 36 unsupported claims, among them release dates printed as event dates and about a dozen unsourced role duties. | A correction pass of 117 edits, then the gate ran again and passed. |
| 21 October | In a skills audit, verifiers refuted or downgraded 33 of 61 claims from the research lanes. | We spot-checked the key refutations in our code and keep a refuted, do-not-re-propose list. |
| 22 to 27 October | We told ourselves Blitz ran at a steady 50,000 to 100,000 records a month. The recount did not reproduce it. | We lead with the 75,397 logged client records over 22 days. |
| 27 October | In the overnight run, 6 of 12 sync steps stopped, because the sync prompt said "only if ready" and was never handed the verdict. The finding text from both verify rounds (4 blocking and 86 structural findings, then 2 and 63) was never saved. | Pass the verdict into the sync prompt. Save findings to disk, because counts alone are not a record. |
| 3 November | The self-hosted search engine returned nothing usable on 117 phrase queries. | Discovery moved to built-in web search. |
Most of these are about verification and wiring. The Blitz row cost us a figure we had said out loud, and our measured piece on BlitzAPI's plans has the full numbers.
Where we put a person
Reads and first drafts go to the agent. Verification goes to a different agent from the one that did the work, and anything that reaches a remote system waits for a person. That means a send, a purchase, a cancellation, a publish or a CRM write.
The line comes from the undo boundary above. A file edit can be rewound, but a sent email, a cancelled mailbox or a charge can't.
An independent verifier with a live fetch caught 5 fabrications on 24 July that the drafting agents' own grounding notes had not flagged. Another refused to call a 30-page dive ready on 15 October. The miss came on 2 September, from a diagram check that gave a prose read at low effort on a mid-tier model and passed three labels that crossed a circle.
The vendor's published server card for the cold-email infrastructure server we drove on 31 October lists 74 tools behind one OAuth scope called "all". Seventeen descriptions are marked paid or destructive, and the card says marked tools need a confirm flag. The tool that sends email isn't one of them. With one scope, whoever holds the connection can call every tool, so the approval has to sit on our side.
The line moves when an action is reversible and ours. Test sends to inboxes we own are the nearest case we have, and even those had a person approving each write. The rule might not bind a team that only sends to its own list, where a bad send has less to break.
Our piece on full autonomy versus human in the loop has the per-action rubric, and our methodology shows where the approval points sit in our research.
A check you can run on any vendor or agency
Lead Line Partners is on this list too, and the same questions work on us.
Frequently asked questions
How does Claude Code work?
Claude Code is an agent that runs in your terminal. It gathers context, takes an action with a tool, checks the result and repeats, chaining many steps. Its tools read and edit files, run shell commands, search and fetch from the web. You can interrupt at any point. File edits can be rewound, but actions on remote systems can't, so a permission mode decides what it may do unasked.
How much does Claude Code cost?
Claude Code is included in every paid Claude plan, as of November 2026: Pro at $17 a month billed annually or $20 monthly, Max at $100, Ultra at $200, and Team seats from $20. It shares one usage pool with chat, and heavy users can switch to pay-as-you-go API credits. Anthropic can change plans and prices at its discretion.
Is Claude Code worth it for GTM work?
In our runs it did the research and first drafts we pointed it at. Whether that pays back depends on a dollar cost per job that we haven't measured, so we can't give you a return figure. Anything that sends, buys or cancels needs a person's approval each time. Try it on ten accounts you already know and compare its work with yours.
What happens when you hit the Claude Code usage limit?
You wait for the window to reset, or keep going on usage credits billed at standard API rates. Most accounts reset on a rolling five-hour window, and paid plans add weekly limits. Terminal work, chat, web and mobile all draw from the same pool. On 4 July our 37-agent run hit the session limit within 4 minutes, so we now pin a model per stage.
Supporting
- Anthropic, Claude pricing, accessed 13 November 2026
- Anthropic, How Claude Code works, accessed 6 November 2026
- Lead Line Partners, run logs, audit records and agent-workflow notes, 9 June to 3 November 2026 (internal records)
Ask in Trellis: Answers From Your Own Workspace, With the Sources Named
Ask opens on the account, person or call you're looking at, answers from your own Trellis workspace, and names the records behind the answer. It doesn't draft or send.
7 Ways AI Research Agents Hallucinate Signals and How to Catch Them
An AI research agent doesn't fail by crashing. It hands back a confident paragraph that's wrong in a way nobody checked. Seven ways that happens, and the check that catches each one.
AI SDR vs ABM: You're Comparing a Channel to a Strategy
An AI SDR automates outreach. ABM decides which accounts deserve it. Comparing the two head to head is why teams end up with a fast channel and no strategy behind it.