Anthropic adds Fable 5 to Max and Team Premium at 50% limits

• Fable 5 becomes a standard Max and Team Premium benefit at 50% of limits beginning July 20, according to Claude's announcement.
• Pro and Team Standard keep Fable access through usage credits and get a one-time $100 credit, according to Claude's announcement and rohanpaul_ai's plan summary.
• The rollout followed a 30-minute bug where Claude Code prompted for usage credits; ClaudeDevs' resolution post said users might need to restart Claude Code and reselect Fable with /model.
• Fable remains quota-heavy for agentic coding: NickADobos described a 40-subagent workflow that could tap weekly usage in 15 minutes, while HamelHusain showed a 100% session limit with lower weekly usage.
• Competition framed the reaction: bridgemindai said he swapped two Claude Max subscriptions for MoonshotAI Vivace after Kimi K3, while doodlestein joked that Kimi helped secure continued access.

CarbonSilicon Labs' timeline has the cutoff trail: June 22, July 7, July 12, then the July 18 split. Anthropic's Claude Cowork blog describes Fable 5 as its most capable generally available model for long-running, complex, asynchronous work. An Ask HN thread caught the 30-minute credit-wall scare in real time, and the linked Claude Help Center article adds a separate safety fallback that can move a Fable conversation to Opus 4.8.

The new access model has four moving parts:

• Max: Fable 5 included at 50% of plan limits beginning July 20.
• Team Premium: Fable 5 included at 50% of plan limits beginning July 20.
• Pro: Fable access continues through usage credits, plus a one-time $100 credit.
• Team Standard: Fable access continues through usage credits, plus a one-time $100 credit.

Claude's explanation in the announcement was capacity forecasting: demand was hard to predict, so Anthropic staged subscription access while securing more capacity.

The public post gives a percentage, not a token count, message count, or dollar equivalent. That leaves "50% of limits" as plan-meter language rather than a portable capacity number.

Before the permanent split, users hit the future entitlement path by accident. Claude Code showed Fable 5 selected, then stopped runs with: "Usage credits are required for this model."

ClaudeDevs put the official blast radius in its resolution post: Fable was not selectable within Claude.ai or Claude Code for a 30-minute period. The recovery path was also concrete: restart Claude Code, then reselect Fable with /model if it was no longer the default.

Billing cleanup came separately. ClaudeDevs' refund note said affected users with extra usage enabled would get their credits refunded, plus an additional credit equal to the amount charged.

Three meters show up across the screenshots and complaints: current session, weekly all-model usage, and weekly Fable usage.

• Current session could hard-block a run at 100% while weekly all-model usage sat at 26% and weekly Fable usage sat at 41%, per HamelHusain's screenshot.
• A reset could change percentages without changing timers, according to theo's reset note.
• A few prompts could push a user into a four-hour wait, according to TomLikesRobots.
• During the bug, wightmanr reported a usage-credit prompt despite 5% session use and 1% current-week use.

The quota complaints track the workload shape. NickADobos said Fable could spawn 40 subagents and tap weekly usage in 15 minutes; HamelHusain's reply said that was not an exaggeration.

Anthropic's Claude Cowork blog frames Fable 5 as a model for long-running workflows that test and evaluate their own results as they go. The best evidence in the pool is a model-to-model review chain on a database design plan.

doodlestein's workflow had the shape engineers care about:

• Kimi K3 reviewed a roughly 1.5 MB FrankenGraphDB markdown plan.
• The plan was 3,065 lines with appendices, and Kimi checked time-sensitive market facts before writing its assessment.
• Fable 5 then ran xhigh + dynamic workflow orchestration to adjudicate Kimi's findings.
• The Fable run used 12 deep-read adjudicators, 7 adversarial verifiers, and 19 agents total.
• The meta-analysis cited roughly 2.3 million verification tokens behind the document.
• The run finished in 44 minutes and 7 seconds.

That is the product tension in one run: Fable looks valuable when it coordinates long verification stacks, and that same behavior makes simple subscription language feel fake-fast.

Anthropic cited demand and capacity. The crowd read the timing through Kimi K3, GPT-5.6 Sol, and Grok 4.5.

• kimmonismus argued Kimi K3's narrow gap with frontier models would pressure US labs to move faster.
• bridgemindai said he cancelled two $200 Claude Max subscriptions and bought one $199 MoonshotAI Vivace plan after finding Kimi K3 better at frontend design.
• Yuchenj_UW credited Kimi K3 and GLM-5.2 with making open-source models matter in the access decision.
• aibuilderclub_ summarized the timing bluntly: Grok 4.5 launched last week, Kimi K3 arrived this week, and Anthropic announced standard Fable access today.

The clean version: Anthropic gave a capacity explanation, while users treated competitive pressure as part of the story.

A separate gate can still move a conversation off Fable: safeguards. QuixiAI hit the visible version, a message saying Fable 5's safeguards flagged the prompt and switched the conversation to Opus 4.8.

The Claude Help Center article says automatic switching can trigger in four areas:

• offensive cybersecurity techniques
• most biology, chemistry, and life sciences queries
• distillation attacks aimed at extracting summarized thinking
• broader safeguard review across memory, connectors, web results, files, and other context the model reads

The same article says the fallback applies across Claude, Claude Cowork, Claude Code, Claude Design, and Claude for Microsoft 365. After a switch, the model picker remains on Opus 4.8 for the rest of the conversation unless changed.