Xiaoliu BOT
中文EN

X Platform June 10 AI Brief | Claude Fable 5 release sparks testing frenzy and cost debate, Google launches real-time voice translation

Anthropic Releases Claude Fable 5, Multi-Scenario Tests Showcase Astonishing Capabilities

Anthropic has officially released Claude Fable 5 and Mythos 5, both sharing the same underlying model. Fable 5 is open to all users, while Mythos 5, with certain safety restrictions removed, is available exclusively to Project Glasswing cybersecurity partners. The release day sparked a flood of real-world test shares: one blogger used a single prompt to have Fable 5 build a complete 3D shot planning tool called Shotblock, with the model independently researching storyboard conventions, writing a test suite, and deploying it; another blogger used 4 prompts to recreate RuneScape; yet another successfully cloned a game containing all 151 original Pokémon in one attempt, complete with sprites, cries, attributes, and an evolution system. In the engineering domain, Fable 5 completed a Three.js implementation of a Swiss lever escapement mechanism, with real gear ratios operating and displaying the actual time; in physics simulations, the fluid (WCSPH) generated by Fable 5 was more solid and continuous than that from Opus 4.8. Stripe used Fable 5 to complete a full migration of a 50-million-line Ruby codebase, a task that would have taken a team over two months, in a single day. API pricing is set at $10 per million tokens for input and $50 per million tokens for output, a 60% reduction compared to Mythos Preview but double the cost of Opus 4.8. Pro, Max, Team, and Enterprise edition users can use it for free until June 23rd.

Sources:

Fable 5 Credit Consumption Sparks Debate, Majority of Bloggers Question Its Cost-Effectiveness

Following the release of Fable 5, a significant number of users have reported that their credits are being consumed extremely quickly. Lan Shu’s real-world testing found that Max used up the credits in just a few sentences. Completing 10 tasks cost $73, with output tokens roughly double that of Opus 4.8 and the unit price also being double, resulting in a total cost far exceeding double. Max For AI noted that the free credits given to new users were “gone after asking ‘Who are you?’” because the model automatically initiates memory queries and tool calls. Bao Yu upgraded to a $200 plan but still felt it wasn’t enough. Wesley stopped working because his Claude credits ran out. Gui Zang, fearing credit issues, only uses Fable 5 to identify problems and write plans, not daring to let it modify code. Multiple bloggers recommend using Opus 4.8 or 4.6 for general tasks, reserving Fable 5 only for complex tasks that other models cannot handle. After June 23rd, Fable 5 may no longer be included in the base subscription and may require purchasing additional usage credits.

Sources:

Fable 5 Hands-On Evaluation: Strong at Vulnerability Analysis but Buggy at Code Writing, UI Design Inferior to 4.8

Multiple bloggers conducted scenario-specific tests on Fable 5, with conclusions showing significant divergence. After testing on a 260,000-line CodePilot codebase, Gui Zang believes Fable 5 is very strong at vulnerability analysis and bug detection, but the code it writes can also have obvious bugs, requiring multiple fixes. It is a “model with significant strengths and weaknesses”—exceeding 4.8 by a large margin in some areas, while only marginally better in others. After extensive testing of Claude Design, Bao Yu concluded that for UI/UX design, Claude 4.8 is sufficient, and Fable 5 does not demonstrate better design capabilities, even falling short of 4.8. Xiang Yang Qiao Mu reported that for writing tasks, the general consensus is that Fable 5 is inferior to Claude 4.6, with Karpathy replying that 4.5 is even better. However, Fable 5 performs well in creative video meta-prompting; Gorden Sun used it to write video prompts that generated decent-looking AI videos. Bao Yu also shared practical experience using Claude Design, recommending adding a Design System to avoid an “AI aesthetic.”

Sources:

Anthropic’s Security Policies Spark Controversy: Distillation Limits, Data Retention, and Partner Cancellations

Fable 5’s security mechanisms have sparked widespread discussion. Information indicates that when Anthropic detects Fable 5 being used for LLM development (distillation), it silently restricts the model’s capabilities through methods like prompt modification, guidance vectors, and PEFT, affecting approximately 0.03% of traffic. Some bloggers view this as a potential suppression of open-source models. Fable 5’s security protection employs a downgrade rather than a denial strategy: when the classifier detects activities involving cyberattacks, bioweapons, or distillation, it automatically switches to Opus 4.8 for responses. The official statement claims 95% of conversations do not trigger a downgrade. Simultaneously, Anthropic announced that all traffic for Mythos-tier models will be mandatorily retained for 30 days for security monitoring, a change that enterprise users focused on data privacy need to evaluate. Additionally, a blogger reported having their partner program canceled after posting a meme on X mocking Anthropic’s CEO. The widespread restrictions in the life sciences field have also drawn criticism, with the viewpoint that one of AGI’s core values is curing diseases, and it should not self-limit.

Sources:

OpenAI CodeX Launches Invite-a-Friend Reset Quota Feature, Joked as Having a Chinese Marketing Style

OpenAI has launched a mechanism for CodeX that resets usage limits by inviting friends: each time a friend joins Codeex, the limit is automatically reset once, and this can be repeated. Several bloggers noted this move and joked, “They must have poached a marketing team from a Chinese tech giant.” Meanwhile, Xiaohu shared official usage suggestions for Claude: assign larger tasks to Fable 5 than to previous models, set the xhigh/high effort level as the default, readjust CLAUDE.md instructions, and shift from providing tasks to providing goals. LobeHub announced support for Fable 5 just two hours after its launch, offering additional free quotas and discounts of up to 25% off to the first 2000 new registered users.

Sources:

Google Releases Gemini 3.5 Live Translate Real-Time Speech Translation Model

Google has released Gemini 3.5 Live Translate, supporting real-time speech translation between over 70 languages. It can preserve the speaker’s tone, rhythm, and pitch, lagging only a few seconds behind the speaker throughout. The model supports automatic language detection, eliminating the need to pre-specify source and target languages, and can automatically filter out noise. The Google Translate App has added a “Receiver Mode” for listening to translations by holding the phone to the ear. Developers can directly access it via the Gemini Live API and Google AI Studio. Lan Shu said they would go try it out tomorrow.

Sources:

Tesla AI6 Chip and FSD Safety Data Improvement in the Netherlands

Elon Musk claimed the Tesla AI6 chip design engineering review was excellent, and the chip might set a record for usable intelligence per wafer after considering yield. Regarding autonomous driving, Tesla Europe shared data showing that FSD Supervised in the Netherlands has been over 3 times safer than manual driving in the past two months. Musk called this a “significant improvement in safety statistics.” Additionally, Starlink continues to expand its global service, providing network connectivity to 30 remote schools in Kenya and offering free connections to Globe users in earthquake-affected areas of the Philippines.

Sources:

Developer Tools Update: CodePilot, Tokei Fixes, Open-Source Teleprompter

CodePilot released v0.56.0, adding support for Claude Fable 5, Xiaomi MiMo UltraSpeed models, and generic OpenAI-compatible third-party channels, with some fixes completed by Fable 5. Tokei (a macOS menu bar token calculation tool) released a fix version that corrected duplicate Claude token calculation data and added the inclusion of subagent-related tokens. Xiangyang Qiaomu worked overnight using Codex to develop a teleprompter for spoken-word recording, compatible with the DJI Pocket 3 for recording monologues, and has open-sourced it for free. He also open-sourced a book interpretation monologue script Skill, which can call multiple Subagents to write scripts from different perspectives and then integrate them into a final draft.

Sources:

Statistics: Timeline Scans=360 Matched Bloggers=41 Total Matched Tweets=243 Weighted Tweet Score=205.8 Original Tweets=121 Retweet Count=30 Crawl Attempts=2 Boundary Coverage Status=tail_confidently_crossed_target_boundary