AI Breaking News is an AI-generated alert, curated and reviewed by the Kursol team. When major AI developments happen, we break down what it means for your business.
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling on July 15—a 975-billion-parameter open-weights model trained on text, image, audio, and video. Unlike GPT-5.6, Claude Opus 4.8, or Gemini 3.5 Pro, Inkling's weights are publicly available, meaning any organisation can download the model, run it on-premise, fine-tune it for proprietary tasks, and avoid paying per-token fees to frontier model vendors. Inkling isn't claimed to be best-in-class—Thinking Machines explicitly positions it as a customisable foundation, not a finished product. But it signals a fundamental shift in how open-source models now challenge the pricing power of closed frontier vendors. For enterprises locked into contracts with OpenAI, Anthropic, or Google, Inkling's release expands your tactical optionality and strengthens your negotiating position.
Murati's Inkling Challenges Closed-Model Economics
Inkling is designed as a mixture-of-experts (MoE) system with 975 billion total parameters, but uses only ~41 billion active parameters per task—a common efficiency pattern that keeps very large models fast and cheap to run. The model supports a 1-million-token context window and was trained on 45 trillion tokens of text, image, audio, and video data. Unlike purely text models, Inkling reasons natively across all four modalities, meaning you can feed images, video, and audio directly without separate encoding or preprocessing steps.
Inkling's core differentiator is that it's downloadable and modifiable. You can host it on your own infrastructure, fine-tune it on proprietary data, and deploy it without paying Thinking Machines a transaction fee. Murati's positioning is explicit: Inkling is a starting point for customisation, not a finished product optimised for general use. Enterprises fine-tune it through Tinker, Thinking Machines' customisation platform, to adapt it for domain-specific tasks—legal document analysis, medical imaging interpretation, supply-chain forecasting, etc.
The release also arrives in a moment of intense frontier-model competition. OpenAI's GPT-5.6 family dropped pricing 50–75% in early July. Google Gemini 3.5 Pro is set to ship with a 2-million-token context window. Anthropic's revenue has leapfrogged OpenAI's. In that environment, an open-weights model from a credible founder gives enterprises a new negotiating lever.
Why Open-Weights Models Undermine Frontier Vendor Pricing Power
For operations and procurement teams, Inkling's release represents a structural shift in your AI vendor alternatives—one that fundamentally changes the economic assumptions you've been operating under.
First: On-premise inference costs collapse when you own the weights. When you pay OpenAI $5 per million tokens for GPT-5.6 Sol, you're paying for model access, compute infrastructure, and support. When you download Inkling and run it on-premise or on commodity cloud infrastructure (AWS, Azure, CoreWeave), your inference cost is just the compute—typically 40–60% cheaper than paying per-token to a vendor. For organisations running millions of inference queries monthly (customer support chatbots, internal knowledge workers, document processing pipelines), this cost differential is material. If your organisation runs significant inference workloads, the question has shifted from "Which frontier model should we use?" to "Can we achieve acceptable quality with an open-weights model and recapture the margin?"
Second: Fine-tuning on proprietary data is no longer vendor-gated. With closed models, you either (a) use OpenAI's fine-tuning API and pay per token, or (b) upload proprietary data to a vendor's infrastructure with data residency risk. With Inkling, you download the weights and fine-tune locally. That means you can adapt Inkling to domain-specific tasks—legal document classification, claims processing, diagnostic imaging—without vendor involvement. The risk that your proprietary training data leaks to the vendor (or gets used to improve the public model) drops to near-zero. For regulated industries (finance, healthcare) where data privacy is non-negotiable, open-weights models just became strategically viable.
Third: Vendor concentration risk is now quantifiable and negotiable. A month ago, if you asked OpenAI "What happens if we want to migrate to an alternative vendor?" the answer was "You'll need to rebuild your integrations and accept some quality degradation." Today, you can run a proof-of-concept with Inkling, Claude, or Gemini, and ask OpenAI: "Here's what we can achieve with an alternative at 40% of our current cost. What's your counter-offer?" That conversation just got real. For the first time, vendor diversification in frontier AI has a credible open-source leg.
What Your Team Should Evaluate This Week
If your organisation is running inference workloads on OpenAI, Anthropic, or Google Cloud at scale:
1. Audit your inference volumes and cost sensitivity by use case. Pull your inference logs from the last 90 days and categorise queries by: (a) tasks where quality/speed is critical and (b) tasks where commodity-grade performance is acceptable. For category (b) workloads, run a 30-day proof-of-concept with Inkling. Document latency, accuracy, and cost. This is your baseline for renegotiating with your current vendor.
2. Benchmark Inkling on your most latency-sensitive workflow. Don't assume Inkling is slow—it's optimised for speed through its MoE architecture. Take your most performance-critical use case (customer support chatbots, real-time document processing) and test Inkling's latency and throughput on commodity infrastructure. Compare to your current vendor's SLA. If performance is within 10% and cost is 50% lower, you've found your renegotiation point.
3. Review your vendor contract for inference price reductions and multi-vendor clauses. Many enterprise contracts with OpenAI or Anthropic have "market conditions" or "competitive pricing review" triggers. If you've been paying premium pricing because alternatives didn't exist, Inkling changes that calculus. Contact your vendor's account team and request a pricing review based on competitive open-weights alternatives. Frame it as a desire to remain with your incumbent, not a threat to migrate—but the threat is implicit.
4. Plan for hybrid inference architecture. Don't assume you'll migrate 100% of inference to a single open-weights model. Design a hybrid approach: frontier models (OpenAI, Anthropic, Google) for high-value, high-accuracy workloads; open-weights models (Inkling, Llama, Mistral) for volume inference where cost matters more than precision. This architecture maximises your negotiating leverage: you can tell each vendor, "We're routing our highest-value work to you, and using open-source alternatives for commodity tasks."
The Bottom Line
Thinking Machines' Inkling release marks the moment when open-source AI models stopped being "good enough for research" and became "cost-competitive for production inference." For enterprises locked into expensive per-token contracts with frontier vendors, this creates immediate optionality: you can now benchmark alternatives, quantify your switching costs, and renegotiate with leverage. The vendors know this. Expect your account teams to proactively offer pricing reviews and longer-term commit discounts before you even ask. The organisations that move fastest on evaluating and benchmarking Inkling against their current workloads will recapture significant budget—and establish the vendor diversity that protects against future price shocks.
If this development has you rethinking your AI vendor strategy, take our free AI readiness assessment to understand where you stand.
AI Breaking News is Kursol's rapid analysis of major artificial intelligence developments—focused on what actually matters for your business. Subscribe to our RSS feed to stay informed.
FAQ
It depends on your use case. Inkling is very capable—it's multimodal, supports long context, and performs well on benchmarks—but it's not universally superior to GPT-5.6 Sol on every task. The real value is cost arbitrage: if you can achieve 85–90% of the quality at 40–50% of the cost, that's a win. Run a 30-day proof-of-concept on one of your highest-volume workloads. If latency and accuracy are acceptable, you've found your renegotiation lever with your current vendor. If they are not, you now have data to optimise your vendor mix.
You can run Inkling on either. A 975B parameter model requires significant compute resources (typically 8 GPUs or equivalent), so on-premise deployment makes sense only if you have existing GPU infrastructure. For most organisations, running Inkling on cloud infrastructure (AWS, Azure, CoreWeave, or Meta Compute) makes more economic sense. You avoid capital expenditure on hardware and get elasticity to scale up or down based on demand. The key point: you're paying compute costs, not vendor transaction fees.
If you have proprietary domain data (contracts, medical images, supply-chain data), fine-tuning Inkling is worth evaluating. Fine-tuning on Inkling means you own the weights after training, so there's no vendor lock-in and no risk of proprietary data being exposed to third parties. For regulated industries (healthcare, finance) where data residency is non-negotiable, this is a major advantage over fine-tuning with OpenAI or Claude. --- If you're uncertain whether your organisation is positioned to adopt open-source models for cost optimisation, [take our free AI readiness assessment](/aiassessment) to evaluate your options.
Ready to get your time back?
No pitch, just a conversation about what Autopilot looks like for your business.