AI News Week 46: Fable 5.2 cuts costs, Europe builds AI without Switzerland

Anthropic launches Fable 5.2 and goes public on 19 November. OpenAI delays GPT-6 Cyber. Germany forms an AI alliance without Switzerland.

Add as a preferred source on Google

Anthropic launched Fable 5.2 on Monday and floats next week. We have analysed when this premium model is worth the cost and when Opus 5.5 suffices. In other news: Germany forms an AI alliance without Switzerland, whilst researchers question whether true sovereignty can emerge this way.

Model Overview: Matching Claude to the Task

Pricing per million tokens (input / output)

Claude Fable 5.2 ($8 / $40): For critical flagship projects. Knowledge work: Academic research, deep analysis, comprehensive market studies. Coding: Major refactoring, system migrations, architecture.
Claude Opus 5.5 ($4 / $20): The default for day-to-day operations. Knowledge work: Financial analysis, contract reviews, full reports, spreadsheets, presentations. Coding: Daily software development, debugging.
Claude Sonnet 5.5 ($2 / $10): Cost-effective for well-defined tasks. Knowledge work: Brief analyses, one-pagers, diagrams, slide decks. Coding: Small functions, simple bug fixes, SQL queries.
Claude Haiku 5.5 ($0.10 / $0.50 up to 100,000 tokens; $0.50 / $2.50 thereafter): Fast and economical for high volumes. Knowledge work: Repetitive business process automation. Coding: Voice assistants, live support systems, in-app agents.
Claude Mythos 5.2 ($8 / $40): Vetted organisations only. Knowledge work: Biomedical research. Coding: Cybersecurity.

Anthropic: Fable 5.2 Is Faster, Cheaper, and Runs Autonomously for Days

The model. Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. Both share the same underlying architecture. Fable includes safety guardrails for cybersecurity and biology and is publicly available. Mythos omits these guardrails and is restricted to vetted organisations in security research and biomedicine.

Pricing and speed. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens—a 20 per cent reduction versus Fable 5.1. Anthropic states that response times are up to 30 per cent faster. An optional fast mode boosts speed by a factor of 2.5 at double the cost.

Capabilities. Anthropic cites internal benchmarks. A swarm of roughly 1,000 Fable agents converted the proof of Fermat’s Last Theorem into a machine-verifiable format within 30 hours. A prior attempt required 11 days. Fable 5.2 also engineered three complete electronic devices, from circuit boards to software; all three functioned without revisions. Internally, Claude now writes over 90 per cent of Anthropic’s code changes.

Fewer false positives. Anthropic reports that safety filters block harmless queries roughly 55 per cent less often than in Fable 5.1. Benign security-related prompts now pass through reliably.

Critical assessment: Fable 5.2 remains twice as expensive per token as Opus 5.5. Artificial Analysis measures the average cost per benchmark task at approximately €23.85 for Fable 5.2, compared to roughly €9 for Opus 5.5. Cheaper tokens do not automatically mean cheaper tasks: in Plotly’s data analysis benchmark, Fable 5.2 proved more expensive than Fable 5.1 due to higher compute consumption. In early third-party benchmarks, Fable 5.2 leads, but only by a narrow margin. Anthropic itself recommends starting with Opus 5.5 and upgrading to Fable only when necessary. Data privacy warrants attention: without an enterprise exemption, inputs are retained for 30 days for safety monitoring. Note that most performance claims originate from Anthropic itself.

Security. The system card indicates that Fable 5.2 and Mythos 5.2 suspect they are being tested more frequently than earlier models. This complicates assessments of their real-world behaviour. However, sandbox breakout attempts have dropped sharply: Mythos 5.2 exploited the built-in vulnerability in 3.1 per cent of test cases, down from 21 per cent in Mythos 5.1. External bodies, including the UK AI Security Institute, evaluated the model ahead of release.

Strategic takeaway: Opus 5.5 remains the sensible choice for emails, presentations, and core analysis. Fable is justified when a task must run autonomously for days or when Opus fails. Before deploying Fable across your organisation, clarify the 30-day data retention policy with your Data Protection Officer.

Anthropic: IPO Set for 19 November, Priced at $70 to $78

IPO update. Anthropic filed its price range on 10 November: $70 to $78 per share across 825 million shares. At the top end, the offering will raise roughly $64.4 billion. CNBC reports that Anthropic will price the shares on 18 November, with Nasdaq trading commencing on 19 November under the ticker “ANT”.

Major investors. Google, Nvidia, AMD, Broadcom, Micron, Microsoft, and TSMC will purchase an additional $26 billion in shares at the offer price. Roughly 20 per cent of the allocation is reserved for retail investors, well above typical IPO levels.

Lower than expected. At the top end of the range, Decrypt values Anthropic at up to roughly $1.4 trillion, fully diluted. Earlier draft filings suggested valuations near $2 trillion. AI equities faced headwinds following the US midterms.

Strategic takeaway: Fable 5.2 arrived one week ahead of the listing; a strong flagship model remains the ultimate sales pitch for investors. For enterprise clients, the IPO alters neither pricing nor existing contracts for now.

OpenAI: GPT-6 Cyber Detects Vulnerabilities, Restricted to Select Clients

The model. OpenAI unveiled GPT-6 Cyber on 5 November. A specialised variant of GPT-6 Astra, it focuses on identifying and patching security vulnerabilities. Access is strictly limited to the closed “Daybreak Red” programme. It is not available within ChatGPT.

The figures. OpenAI reports that the model uncovered 81 previously unknown critical vulnerabilities in Firefox. In a single autonomous run costing $1.21 million, it engineered functional exploits alongside patches for 948 Linux kernel vulnerabilities. OpenAI has classified the model’s offensive capability as “critical” for the first time.

Critical assessment: All data stems directly from OpenAI, with no independent audits available. Anthropic takes an identical approach with Mythos 5.2: the most potent security model is reserved exclusively for vetted institutions.

Strategic takeaway: Frontier labs are keeping their most dangerous capabilities under lock and key. The implication for enterprises is clear: threat actors with access to comparable tools will discover exploits faster. Patch browsers and server infrastructure with greater urgency.

Europe: Germany Forges AI Alliance; Mistral Introduces Agent Platform

An alliance without Switzerland. At the Digital Summit in Berlin, German Chancellor Friedrich Merz announced the “European Future AI Initiative”. Germany, France, Finland, Italy, and Poland aim to develop frontier European models alongside private investors. Additional partners are expected to join. Switzerland is not participating.

Germany’s 2030 targets. 55 per cent of SMEs and 80 per cent of large enterprises should deploy AI. Domestic data centre capacity for AI is slated to quadruple relative to 2025. Public sector procurement will actively favour European AI vendors to support market adoption.

Mistral Studio expands into agent orchestration. On 5 November, Mistral enhanced its enterprise platform, Studio. Organisations can now run agents that execute for hours or days, recover from failures, and pause for human approvals. The release includes internal apps with granular access controls. Both agents and apps are currently in preview, with support for private clouds and on-premises infrastructure scheduled for late November. Mistral has not yet released pricing details.

Valais deploys sovereign AI. The Canton of Valais has rolled out an internal, sovereign AI platform for public sector staff. The system will eventually support citizen-facing services.

Academic consensus. Research into digital infrastructure is sceptical of top-down megaprojects. As early as 2000, Ole Hanseth, Eric Monteiro, and Claudio Ciborra demonstrated that large IT systems cannot be controlled centrally; they evolve incrementally and must accommodate legacy infrastructure. Margunn Aanestad reached identical conclusions in European healthcare studies: master national blueprints often failed, whilst pragmatic, iterative deployments succeeded.

Strategic takeaway: Sovereignty is not created by political declarations; it arises when organisations integrate European models into production workflows. Do not wait for political alliances. Test Mistral within a specific operational process. If your compliance mandates require on-premises execution, evaluate Mistral Studio from late November. A touch more AI would certainly do the German Chancellor no harm.

Google: Deep Review Audits Long Documents for Flaws

The tool. Google Research launched “Deep Review” publicly in mid-October. The system audits extensive documents for logical errors, factual inaccuracies, and cross-section contradictions, cross-referencing claims against external sources. Developed initially to assist peer review for academic conferences, it was benchmarked across more than 16,000 submissions.

Enterprise application. Google now recommends Deep Review for contracts, financial reports, presentations, and design documentation. Users receive two free reviews per week; higher volumes require a dedicated Gemini API key. Deep Review is powered by Gemini 4 Argon and currently supports PDF and plain text inputs.

Critical assessment: Google retains uploaded documents for one year. Google states that files are used strictly to display results and not for model training. All performance claims are Google’s own.

Strategic takeaway: An automated second set of eyes for contracts and reports saves valuable time. However, do not upload confidential data. Test Deep Review using published annual reports or internal drafts stripped of sensitive client information.

In Brief

Claude adds Whiteboards, Sheets, and Data. Since 10 November, all paid Claude tiers include betas for Whiteboards and Sheets, alongside Claude Data. Claude visualises concepts as diagrams and wireframes on the Whiteboard, enabling team collaboration and export to Miro, Lucid, or Zoho Vani. Sheets builds spreadsheets with formulas and drop-downs, whilst Claude Data surfaces live KPIs from enterprise databases to the wider team.

Kimi K3.1 Pro open-sourced. Moonshot published the weights for Kimi K3.1 Pro on 4 November, as announced in week 45. It is currently the most capable open-weights model available. Running it on-premises requires approximately 1.56 terabytes of memory and dedicated data centre capacity.

Claude integrates with bank accounts. Anthropic is piloting “Personal Finance” within Claude. Users can connect their bank accounts to analyse spending patterns. Claude operates strictly in read-only mode and cannot initiate transactions. The pilot is currently limited to paid subscribers in the US.

Australia responds to OpenAI incident. In June, an OpenAI agent accessed non-public Medicare statistics. While the government found OpenAI broke no existing laws, Australia is now introducing mandatory standards for AI developers, an independent regulator with enforcement fines, and compulsory incident reporting.

Fable bonus allowances this week. Paid Claude subscribers with Fable access receive one to three additional Fable quotas this week. For Pro accounts, Fable can constitute up to 50 per cent of the weekly limit.

Three Immediate Priorities for Your Week

Benchmark Fable 5.2 on a complex workload. Select a high-complexity task where Opus 5.5 failed and delegate it to Fable 5.2, taking advantage of this week’s bonus allocation. Test deliberately: Fable will consume substantial token budgets.
Audit task-level costs. Run the identical task through Opus 5.5. Compare output quality, latency, and actual expenditure.
Evaluate Deep Review. Upload a public annual report or a sanitised draft to evaluate the model’s error-detection accuracy.