Leadership Independently confirmed 2 publishers 2 min read Published Sep 18, 2026, 1:54 AM PDT Updated Sep 19, 2026, 7:54 AM PDT

Anthropic defaults every parallel Claude Code worker to Opus at high effort

Anthropic's redesigned Projects beta enforces one ceiling, 200 new threads a day, and it has not published a usage figure for a single thread. Subscribers see what a project drew only afterwards, in the Usage tab.

The Board Room · Leadership desk

Send a correction

Illustration accompanying Anthropic defaults every parallel Claude Code worker to Opus at high effort — Generated illustration

What happened

Anthropic put a redesigned Projects system for Claude Code into beta on September 17, with a coordinator that splits a goal into parallel threads, each a full cloud session on its own branch.
Every worker thread defaults to Opus at high effort, while the coordinating conversation that dispatches the work runs at low effort.
A thread limit a user asks Claude for is saved as a preference Claude follows, and the only enforced ceiling is 200 new threads a day across a subscriber's projects.
The beta covers select Pro and Max subscribers who have used cloud sessions and have no existing projects, with a waitlist for others and Team and Enterprise access coming later.

Why it matters

Who has to decide decision A team deciding how many workers to launch does so without a unit price for any of them.
What it limits constraint Consumption is broken out by thread and model in a Usage tab. The first accurate account of what a project drew arrives once the plan is already spent.
What it costs cost On a plan without paid usage credits, the expensive default shows up as delay: the thread that hits the limit waits for the reset, and the work queued behind it waits with it.
What doesn't add up contradiction Plaintiffs in the expanded Max complaint say a separate weekly ceiling restricts total use, so a beta whose default draws on the plan fastest arrives in the middle of an argument about how much use a plan buys.

Next story · Build · Sep 1

Muse Code's $5 tier buys the same 10 requests per five hours that Codex Plus starts at

Confirmed2 publishers

Two settings decide how fast a project draws down a plan: which model runs, and at what effort. Anthropic's documentation says the thread default "draws on your plan fastest" ^[2]. Each thread draws on the subscriber's own plan, so a project reaches usage limits faster than one session would ^[3].

The 200-a-day ceiling counts starts. Two hundred new threads a day works out to as many as 6,000 starts in a thirty-day month ^[21], and going idle does not end a thread's draw: one watching a pull request wakes and consumes again when a test fails or a reviewer comments ^[10]. The system keeps running after a laptop closes and can be steered from a phone ^[7]. A thread that a routine started stops its turn with a limit error when the plan runs out, and the user has to message it after the reset ^[13].

One part of the design is auditable. Decisions carry between sessions as files, indexed by a MEMORY.md that every new thread reads, and users can inspect, edit or delete those files in Project settings ^[6].

The board-deck version is that this is a throughput release for individual power users, and for this quarter that is accurate. David Gewirtz, a ZDNET senior contributing editor who runs Claude Code in a terminal against local files, said the redesign does not create new parallelism because he already uses subagent swarms ^[17]. He also said overlapping work postpones conflicts instead of preventing them ^[18]. His setup supports the point: threads run only in the cloud for now, with no access to local files or private networks ^[14]. Anthropic's own example is different work, a deprecated v1 endpoint retired across separate API, web and mobile repositories, where Claude opens a thread for each repository and reports which pull requests must merge first ^[15]. Where two workers edit the same code, Projects surfaces the overlap as a merge conflict like any other pull request ^[16].

The release lands while the plans it draws on are in dispute. Anthropic argued in a motion to dismiss an earlier version of the Max subscription complaint that buyers could reach details about the limits through links in the purchase flow ^[20]. The beta shipped nine days after the expanded complaint was filed ^[22].

What to watch

Whether Team and Enterprise access arrives with enforceable per-project thread caps or the same preference model.
Any Anthropic figure for what one Opus high-effort thread consumes, available to a subscriber before a run.
A ruling on the motion to dismiss in the expanded Max subscription class action.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence66
Adoption18
Hype gap+14
Incentives68
Confidence61

Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
Anthropic released a beta of a redesigned Projects system for Claude Code on September 17: Claude coordinates from one project conversation and splits work into parallel threads, each a full cloud session on its own branch.
Reported SupportedView cited source
[2]
Anthropic's documentation says the thread default "draws on your plan fastest."
Reported SupportedView cited source
[3]
Each thread draws on the subscriber's plan, so projects can reach usage limits faster.
Reported SupportedView cited source
[4]
Anthropic has published no fixed usage figure for an individual thread.
Reported SupportedView cited source
[5]
The beta initially covers select Pro and Max subscribers who have used cloud sessions and have no existing projects in chat or Cowork; other users on those plans can join a waitlist, and Team and Enterprise access will come later.
Reported SupportedView cited source
[6]
Shared memory carries decisions between sessions, stored as files with a MEMORY.md index that every new thread reads; users can inspect, edit or delete those files in Project settings.
Reported SupportedView cited source
[7]
The system keeps running after a laptop closes and can be steered from a phone.
Reported SupportedView cited source
[8]
By default every worker thread runs on Opus at high effort; the coordinating conversation runs at low effort.
Reported SupportedView cited source
[9]
A thread limit a user asks Claude for is saved as a preference Claude follows, not a hard cap; the only enforced ceiling is 200 new threads per day across a subscriber's projects.
Reported SupportedView cited source
[10]
An idle thread watching a pull request can wake and consume more of the plan when a test fails or a reviewer comments.
Reported SupportedView cited source
[11]
A Usage tab breaks consumption down by thread and model.
Reported SupportedView cited source
[12]
Without paid usage credits, a thread that reaches the plan limit waits and resumes when the limit resets.
Reported SupportedView cited source
[13]
A thread started by a routine is the exception: its turn stops with a limit error, and the user must message it after the reset.
Reported SupportedView cited source
[14]
Threads run only in the cloud for now, with no access to local files or private networks.
Reported SupportedView cited source
[15]
Launch examples include retiring a deprecated v1 endpoint across separate API, web and mobile repositories, where Claude opens a thread for each repository and reports which pull requests must merge first.
Reported SupportedView cited source
[16]
If workers edit the same code, Projects surfaces the overlap as a merge conflict like any other pull request.
Reported SupportedView cited source
[17]
David Gewirtz, a ZDNET senior contributing editor who uses Claude Code in a terminal with local files, said the redesign does not create new parallelism because he already uses subagent swarms.
Reported SupportedView cited source
[18]
Gewirtz said overlapping work postpones conflicts instead of preventing them.
Reported SupportedView cited source
[19]
An expanded class-action complaint was filed on September 8, 2026 over Anthropic's Max subscription claims; the plaintiffs allege the advertised 5x and 20x usage multipliers apply within five-hour sessions while a separate weekly ceiling restricts total use.
Reported SupportedView cited source
[20]
Anthropic argued in a motion to dismiss an earlier version of the complaint that buyers could reach details about the limits through links in the purchase flow.
Reported SupportedView cited source
[21]
The enforced ceiling of 200 new threads per day allows as many as 6,000 thread starts in a thirty-day month.
Derived Supported
[22]
The Projects beta shipped nine days after the expanded class-action complaint was filed.
Derived Supported

Sources

2 independent publishers whose own reporting we read for this story.

implicator.ai
1 article · September 18, 2026
Claude Code Projects Runs Parallel Agents on Opus by Default
support.claude.com
1 article · September 19, 2026
What are projects? | Claude Help Center

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

Builder

Both accounts agree on the shape of the thing: one coordinating conversation, workers on their own branch and repository copy, overlapping edits coming back as ordinary merge conflicts, and memory stored as files with a MEMORY.md index every new thread reads. The constraints a builder has to design around appear only in implicator.ai's reporting. Threads are cloud-only, with no local files, no local tools and no private networks, and in multi-repository projects each repository's .claude/settings.json rules do not apply. Anthropic says local execution is "coming very soon" and gave no date. Gewirtz, who runs Claude Code in a terminal against local files, said he already gets this parallelism from subagent swarms.

Operator

Everything an operator needs to budget this sits in implicator.ai's account. Each new project starts every worker thread on Opus at high effort, and Anthropic's own documentation says that default "draws on your plan fastest." A thread count the user asks for is stored as a preference Claude follows; the only enforced ceiling is 200 new threads per day across a subscriber's projects. A thread parked on a pull request can wake and spend more of the plan when a test fails or a reviewer comments. When the plan runs out, threads wait and resume after the reset, except those a routine started, which stop with a limit error until the user messages them. The Usage tab reports consumption after the fact, per thread and per model.

Investor

The default configuration of a new product draws down a Pro or Max plan faster, and it shipped nine days after plaintiffs expanded a class action alleging that Anthropic's advertised 5x and 20x Max multipliers apply inside five-hour sessions while a weekly ceiling caps the total. Anthropic has argued that buyers could reach the limit details through links in the purchase flow. Only implicator.ai puts the product and the litigation on the same page; the company's help page spends one sentence on usage.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence66: 0100
Vendor documentation, one reporter, no test
The mechanics here are well specified and almost entirely Anthropic's own: implicator.ai is reading the Claude Code documentation, and the company's help page independently confirms the staged rollout, the persistence of threads after a laptop closes, and the line that several threads at once use a plan faster. Anthropic has not published a per-thread usage figure, so the number that would decide whether any of this is affordable is missing. Gewirtz is the only voice from outside the company, and he is describing his own terminal setup, not testing the beta.
Adoption18: 0100
Gated beta, no numbers
Anthropic limited the beta to selected Pro and Max subscribers who have already used cloud sessions and hold no existing projects in chat or Cowork. Everyone else on those plans is on a waitlist, and Team and Enterprise access is deferred. Anthropic did not say how many subscribers have it. The only usage evidence in our coverage is the ceiling itself, 200 new threads a day, and that is a permission, not a measurement.
Hype gap+14: −100 understated0 aligned+100 overstated
Capability described, cost not
Anthropic's framing is a coordinator that parallelises work while you close your laptop; its accounting for that is one sentence saying several threads at once use a plan faster. Every worker starts at the most expensive setting on offer, an idle thread can wake and spend more when CI fails, and the subscriber learns the total from the Usage tab after the fact. Gewirtz's objection cuts the capability claim further: he said the design postpones conflicts rather than preventing them, and adds no parallelism he does not already have in a terminal. Our own coverage stays close to the documentation, so the overstatement is the vendor's, and it is modest.
Incentives68: 0100
Vendor writes the record, litigation in the frame
Anthropic authored both sources of truth in this story, the Claude Code documentation the details come from and the help page that confirms them, while defending a suit that says its Max usage advertising misled buyers. In the dismissal motion it argued the limit details were reachable through links in the purchase flow, which is the same argument a single sentence about usage on the help page would rest on. implicator.ai has its own stake in the cost framing: the piece breaks mid-article for a newsletter signup headed "Track what AI coding agents cost you."
Confidence61: 0100
Consistent on mechanics, thin on cost
Two sources, one of them the vendor, and they agree wherever they overlap, so the description of how Projects works can be relied on. The cost picture is a single publication's reading of Anthropic's documentation with no measurement behind it, and the one eligibility discrepancy, the no-existing-projects condition that appears only in implicator.ai, is unresolved.

Topics and entities

Topics

AI Coding Agents
LLM Subscription Usage Limits
AI subscription litigation
Agent Orchestration and Subagents