A week about price and reach rather than the frontier: the cheapest capable Claude yet, Europe's largest open-weight model, AI moving into the tools people already use, and a Croatian battery maker entering AI infrastructure.
Last week was about the frontier: Gemini 4 Argon, Sonnet 5.5, GPT-6.1 Sol and the Astra model OpenAI chose not to ship. This week was about price and reach. Anthropic shipped Claude Haiku 5.5 at $0.10 per million input tokens, a tenth of its predecessor's list price. OpenAI made GPT-6 the default for every ChatGPT user. Mistral answered from Paris with Mistral Large 4, a one-trillion-parameter model whose weights it promises to release this month. Money kept flowing: DeepSeek is reportedly raising around $12B, and Manus raised $500M+ after its Meta deal was unwound. At home, Rimac Technology moved into AI data-centre power systems and Croatian fintech Farseer closed a $7.2M Series A.
The Releases at a Glance
$0.10 / $0.50 per million tokens for prompts up to 100K tokens; $0.50 / $2.50 above that. 1M-token context, 128K output, adaptive thinking with effort control. Available on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry.
1T total / 52B active parameters (mixture-of-experts), natively multimodal, 160+ training languages. Public preview at $1.36 / $4.18. Open weights promised by the end of October. Trained on European infrastructure.
GPT-6 is rolling out globally to all ChatGPT users with "Intelligent UI": answers that include visuals and interactive elements you can use directly in the reply.
Generally available image model with better mask-based editing, subject consistency and text rendering. Output up to 4K, up to 14 reference images per prompt. Live in the Gemini app, AI Studio, Google Ads and Gemini Enterprise.
Claude Haiku 5.5: Read the Price Table Carefully
Haiku 5.5 is the most aggressive price move of the year. Haiku 4.5 costs $1 / $5 per million tokens; Haiku 5.5 costs $0.10 / $0.50, with cache hits at $0.01 and the Batch API halving it again to $0.05 / $0.25. It also brings the 5.5 generation's 1M context window and adaptive thinking to the small tier. For classification, extraction, routing, document triage and customer-support first drafts, this is now the default to test.
Two details in Anthropic's documentation change the maths, and most headlines skip them:
- A price cliff at 100K tokens. If a request's prompt exceeds 100,000 tokens (cache reads and writes included), the whole request is billed at $0.50 / $2.50, five times the short-prompt rate. Long-document workloads still save money compared with Haiku 4.5, but far less than the headline suggests.
- A new tokenizer. Claude 4.7 and later models, Haiku 5.5 included, use a tokenizer that produces roughly 30% more tokens for the same text than Haiku 4.5's.
Worked example: a prompt that counts ≈10K tokens on Haiku 4.5 with a 1K-token answer costs about $0.015 on Haiku 4.5. On Haiku 5.5 the same text is ≈13K tokens in and ≈1.3K out: about $0.002, roughly 87% cheaper. A 100K-token contract review (≈130K on the new tokenizer) crosses the cliff and lands at about $0.07 versus $0.10, only ~30–35% cheaper. Practical rule: chunk long documents below 100K tokens per request, or send them to Sonnet 5.5, which has no long-context surcharge.
Anthropic also flags breaking changes for anyone migrating from Haiku 4.5: manual extended thinking (budget_tokens) now returns a 400 error, and adaptive thinking is on by default, so responses can start with thinking blocks. Test your parsers before switching the model ID. The same day, prompt-cache reads on Sonnet 5.5 dropped from $0.20 to $0.10 per million tokens. Agent workloads that re-send long system prompts and tool definitions on every turn get a direct discount.
Anthropic Pushes Into the Workplace — and Into Startups
Claude for Google Workspace enters public beta
Claude now opens in a sidebar inside Google Docs, Sheets and Slides. It can read the open file, see what you have selected and edit in place: restyle a document, build a financial model in Sheets, lay out slides from your template. Edits wait for approval by default; an "accept all" mode is available. New Docs, Sheets and Slides connectors also let Claude create and edit Google files from the Claude app. Admins deploy it from the Google Admin console; on Team and Enterprise plans an owner must enable the connectors. For the many Croatian SMEs that run on Google Workspace rather than Microsoft 365, this is the first serious alternative to Gemini inside their own documents.
Claude for Startups: a free year of Claude Team plus $1,000 in API credits
Launched during SF Tech Week, the expanded programme gives eligible startups a free year of Claude Team (up to five premium seats), $1,000 in API credits, Claude Marketplace access and office hours with Anthropic's Applied AI team. Eligibility: companies founded in the last five years or funded in the last two. Croatian founders qualify on the same terms. Apply at claude.com/programs/startups.
Claude Frontier Academy: $100M to train 10,000 deployment engineers
Announced on 2 October, the academy aims to train 10,000 "Frontier Deployed Engineers" by the end of 2027. These are engineers at partner firms who take Claude from prototype into production at client organisations. First cohorts in San Francisco, New York and London include engineers from Accenture, Bain, Capgemini, Deloitte, McKinsey and Morgan Stanley, with a 12-week residency leading a real Claude project. The message for buyers: the bottleneck is no longer the model but people who can deploy it.
Smaller items worth knowing: Claude Max and Team plans now include monthly API credits; Barclays is scaling Claude across operations and client service; Anthropic expanded its Cyber Verification Program, which gives vetted security professionals access to advanced cyber capabilities; and it published a 2026 Usage Policy update on 8 October. If you build on the API, read the policy update before your next compliance review.
OpenAI: GPT-6 for Everyone, Atlassian, and Mathematics
GPT-6 becomes the ChatGPT default, with Intelligent UI
A week after DevDay, OpenAI rolled GPT-6 out to all ChatGPT users. Answers can now include visuals and interactive elements. For companies this matters mostly through shadow AI: your employees' free personal accounts just got a frontier model. If your AI policy assumes the free tier is "the weak one", update it.
Atlassian brings GPT-6 models into Jira and Confluence
OpenAI and Atlassian expanded their partnership to bring GPT-6-family models across Atlassian's platform for planning, building and delivering work. Together with Claude in Google Workspace and Copilot in Microsoft 365, this confirms the 2026 pattern: AI is reaching most employees through the tools they already use, not through a separate chat window.
New results on open mathematics problems, with Lean proofs
OpenAI published results on open mathematical problems produced by an internal frontier model and released Lean formalisations of the proofs on GitHub. The significant part is not the headline but the method: machine-checkable proofs. The business analogue is pairing a model with a verifier, such as tests for code, reconciliations for finance or schema checks for data, rather than trusting fluent output.
Also reported this week (WSJ): Broadcom is seeking more than $50B in financing tied to OpenAI's custom "Nexus" chips. The compute race we covered in Wednesday's infrastructure article is now being financed with debt at sovereign scale.
Europe: Mistral Large 4 and the Sovereign Option
Mistral's first major launch since its €3B Series D is its biggest model to date: a 1T-parameter mixture-of-experts with 52B active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres. Mistral reports top-five placement on the Artificial Analysis Cyber Index, 93% on Cybench and 61.7% on DeepSWE v1.1. On coding it trails the US frontier but competes. At $1.36 / $4.18 in public preview it undercuts the $2 / $10 frontier tier. Mistral offers a European deployment it operates end to end, as well as private-cloud and on-premise options.
Why it matters for Croatian companies: open weights at this scale mean banks, insurers, public institutions and healthcare providers can run a near-frontier model on infrastructure they control. With 160+ training languages, Croatian is a realistic target. Our advice is the same as for every release: build a 50–100-task test set in Croatian from your real documents and measure before you commit. Weights are due by the end of October.
On the regulatory side, nothing new took effect this week, but the clock is running. Under the Digital Omnibus, generative AI systems already on the EU market must implement machine-readable output marking by 2 December 2026. The Microsoft 365 Copilot licensing change below has an even earlier date.
Usage-based billing on by default for new Copilot Business subscriptions
From 1 December 2026, usage-based billing will be enabled by default for new Microsoft 365 Copilot Business subscriptions bought through CSP partners, the route most Croatian SMEs use. Copilot Search on Windows and the web also gained conversational chat. If you are budgeting Copilot for 2027, ask your partner how metered usage will appear on your invoice and set spending caps.
Funding: Agents and Asian Labs
DeepSeek ~$12B, Manus $500M+, EliseAI $350M, Arena $200M
Bloomberg and Reuters report DeepSeek is close to raising at least $12B (≈80B yuan) backed by Tencent and CATL, ahead of a possible 2027 IPO; the company has not confirmed. Manus raised over $500M led by Boyu Capital and IDG, its first round since Chinese authorities forced the unwinding of Meta's $2B acquisition; it recently launched Manus 2.0 and Cue, which gives agents their own email, phone numbers and wallets. Property-management AI EliseAI raised $350M at $4B. Arena, the model leaderboard, raised $200M at $3.1B. That valuation says something: independent evaluation has become a business in its own right.
Tutorial of the Week: Route Cheap, Escalate Smart
Haiku 5.5's pricing makes a classic pattern almost free to run: a router-and-escalation pipeline.
- Triage with the small model. Send every incoming item (email, ticket, invoice, CV) to Haiku 5.5 with a strict JSON schema: category, urgency, language, extracted fields and a confidence field.
- Escalate on rules, not on vibes. Forward to Sonnet 5.5 or Opus 5.5 only when confidence is low, the item exceeds a monetary threshold, or the category is on a "sensitive" list (legal, HR, complaints).
- Keep prompts short. Stay under the 100K-token cliff by sending relevant excerpts, not whole archives. Cache the system prompt and tool definitions.
- Batch what can wait. Overnight reclassification and back-office extraction belong in the Batch API at half price.
- Measure the escalation rate weekly. If more than 20–30% of items escalate, improve the triage prompt or examples before upgrading the model.
One security note from this week's platform release notes: in Claude Managed Agents, web_fetch now only fetches URLs that already appeared in the session. This blocks a common prompt-injection exfiltration route, where a malicious document tells the agent to "visit" an attacker URL with your data in the query string. Apply the same rule to your own agents: no fetching of model-invented URLs.
Croatia Corner: Batteries for AI, a Series A and a Rijeka Meet-up
Rimac Technology and ECOBLOX build a modular AI data centre
Rimac Technology signed an agreement with US company ECOBLOX to co-develop a modular data centre for AI. Rimac's automotive-grade UNI battery modules, battery management and power electronics will be combined with ECOBLOX's GPU computing, hybrid air and liquid cooling and management software into one unit covering power conversion, UPS, cooling and IT. A demonstration system will be installed at Rimac Campus near Zagreb in November. This is the most concrete sign yet that Croatian industrial know-how can supply the AI infrastructure boom, even while Croatia still has no EU AI factory of its own.
Farseer closes a $7.2M Series A
Croatian AI-native planning and analysis platform Farseer, which competes with Anaplan, Pigment and Excel, raised a $7.2M Series A led by Aymo Ventures, with SQ Capital and US-based Apertu Capital. It now has about 60 employees and customers in 15 European countries, and is expanding across Europe and the US. At the Finance Weekend conference in Rovinj, co-founder Luka Mijatović warned that AI makes it "incredibly easy" for finance teams to get lazy. Hrvatski Telekom's CFO Matija Kovačević said AI made HT's hunt for "ghost payments" 90% more efficient, and expects value in finance to shift "from production towards interpretation".
"AI pogon za biznis" brings entrepreneurs to Veleučilište PAR
On 7 October, the entrepreneurs' association Glas poduzetnika ran a free, practical AI conference in Rijeka. Topics included AI agents in practice (costs, human oversight, and "when not acting is cheaper"), AI-generated fraud, data leaks through public tools, and a closing panel on the Rijeka–Istria AI ecosystem. Demand for practical AI knowledge beyond Zagreb is real, and regional events like this deserve more support.
Two practical notes for Croatian readers. First, with Haiku 5.5 and Mistral Large 4 both supporting Croatian, the cost of testing AI on Croatian-language business documents fell sharply this week. Second, the Claude for Startups programme is open to Croatian companies on the same terms as US ones. A free year of Claude Team is a low-risk way for an early-stage team to build AI into its workflow from day one.
The Week in Numbers
What to Watch
- Mistral Large 4 weights: due by the end of October. Check the licence terms and the hardware needed before planning an on-premise pilot.
- Rimac Campus demo in November: the first look at how an automotive battery supplier positions itself in AI infrastructure.
- Gemini 4 Argon general availability: still limited to vetted defenders. Watch for Vertex AI pricing and EU-region terms.
- 1 December: metered billing by default for new Microsoft 365 Copilot Business subscriptions through CSP.
- 2 December: AI Act machine-readable marking deadline for generative AI systems already on the market.
Haiku 5.5, Sonnet 5.5 or Mistral Large 4 — what does your workload actually need?
Model prices dropped again this week, but only for workloads designed around the price table. We build routing pipelines, evaluate models on your Croatian-language documents and set up agents with tight permissions and human checkpoints. Claude Certified Architect, based in Zagreb.
Talk to us