July 7: Fable 5 is still online!
Just before midnight, Anthropic unexpectedly announced that the strongest Claude Fable 5 will be available for free until July 12.
Five additional days of free access beyond the original plan.
The usage is exactly the same as before: weekly usage is limited to 50%. If you exceed this limit and wish to continue using the service, you must purchase credits.
The wish made by the netizen has finally come true.
The entire network is abuzz. Everyone is spreading the word: What are you waiting for? Get in now!
Squeeze every last bit out of Fable 5—five days left.
Just yesterday, just as everyone believed the shutdown of Fable 5 was inevitable, a wave of "rescue" guides swept through the community.
Developer Alex Prompter immediately pulled off a miraculous feat by perfectly porting Fable 5 to Opus 4.8.
Although there is an additional five-day "probation period," these hands-on self-help methods remain highly effective.
He provided a set of prompts to extract Fable 5's "way of thinking" and turn it into an executable "manual."
Next, use the Project command or API system prompt to migrate the manual into Opus 4.8;
Finally, use a trick question—“5% is said to be 20%”—to verify whether the migration has truly taken effect.
Another developer, Machina, outlined five steps to extract maximum value from Fable 5, including:
- 1. Have Fable rewrite CLAUDE.md and the skills;
- 2. Conduct advisory-style business audits;
- 3. Dig deep research into an atomized Obsidian knowledge base;
- 4. Maximize unattended time using /goal and dynamic workflows;
- 5. Install another skill that automatically records "what it thinks."
The core logic is simple: extract over 50,000 answers from state-of-the-art models to train a smaller model, at a cost of less than $500—the teacher retires, but the student lives on forever.
Anthropic unveils cost-saving strategies
46% cost, 92% power increase
Just as people were lamenting how quickly Fable 5 was burning through funds, Anthropic officially stepped in and unveiled two architectures—
- Advisor mode
- Orchestrator pattern
In simple terms, it’s letting the large model do the talking and the small model do the legwork.
First strategy: Advisory model.
The primary executor is Sonnet 5; it only knocks on Fable 5’s door to ask for advice when it encounters critical uncertain points.
The vast majority of tokens are settled at the low fee rate of Sonnet 5.
On the rigorous code testing benchmark SWE-bench Pro, the results are impressive—
Pairing Sonnet 5 with Fable 5 as an advisor achieved approximately 92% of Fable 5’s standalone performance at just 63% of the cost.
Fable 5 is typically called upon only once per task to guide the direction at decision points, while Sonnet 5 handles all the remaining heavy lifting.
This is exactly what it means to get big results with a small investment.
Anthropic has documented this advisory model in its official documentation and provided a dedicated advisor tool, making it easy for anyone to implement.
The logic isn't complicated: in any task, there are very few instances that actually require "high-level judgment"; most of the time, it's just mechanical reading and writing.
Let your model act only at a few key moments, and your bills will naturally decrease.
Second tactic: the Orchestrator pattern.
Under this mode, Fable 5 is directly promoted to commander, strategically planning and dividing tasks to assign work to the underlying Sonnet 5 sub-agents;
All token-intensive research work is to be pushed down.
On BrowseComp, where massive reading is required, this combination delivers the final blow—
Fable 5's commander achieved 96% of the performance of a single model, while reducing costs to an astonishing 46%.
Looking at the ratio alone isn't very intuitive, so Anthropic directly shared a real bill in their cookbook.
The task is to verify the ticketing and reservation policies for the 10 largest national parks in the United States, with 20 facts, each one cross-checked individually against the official nps.gov pages—no one is allowed to answer based on memory or guesswork.
The characteristic of this task is “reading a lot”: just feeding dozens of web pages into the model results in the majority of the cost being spent on reading.
Address: https://github.com/anthropics/claude-cookbooks/blob/main/managed_agents/CMA_plan_big_execute_small.ipynb
Thus, the value of specialization becomes apparent—
Commander Fable 5 is responsible for planning and aggregating, and doesn't touch a single webpage;
Sonnet 5 sub-agents concurrently read web pages within their respective contexts and report back the distilled conclusions.
Under the same verification standards, the bill comes to approximately $1.61 for the team-based approach; switching to a single Fable 5 tackling all 20 facts directly spikes to approximately $4.
The team not only reduced costs by approximately 2.5 times but also sped up processing by three times—194 seconds versus 608 seconds—with over 80% of tokens allocated to the lower-cost worker rates.
When you look at the two moves together, you’ll see that Anthropic is actually sending the same signal:
The correct use of a top-tier model is never to run it at full capacity all the time.
The free window for Fable 5 closes on July 12.
But Anthropic’s real bombshell over these past few days wasn’t “using the strongest model for free”—it was showing you step by step how to keep using it affordably even after the window closes.
Right now, the most cost-effective approach: Use Fable 5 as your commander and pair it with the affordable Sonnet 5 as your worker!
Reference materials:
https://x.com/claudeai/status/2074548242386178258
https://support.claude.com/en/articles/15424964-claude-fable-5-promotional-access
https://x.com/ClaudeDevs/status/2074606058128224365?s=20
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation, edited by Peach.