Daily AI brief · July 17, 2026
Thinking Machines' Open-Weight Model Challenges Frontier Pricing Power
Thinking Machines released Inkling, its open-weight model with 975 billion total parameters activating 41 billion per task, designed for teams to customize themselves
Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling, its first in-house AI model .
Unlike OpenAI, Anthropic, or Google's flagship models, it's open-weight, and it's a mixture-of-experts system with 975 billion total parameters, activating 41 billion per task .
In a project with Bridgewater Associates, an open-source model trained on financial expertise scored 84.7% on financial reasoning tests, beating top proprietary models while costing roughly a fourteenth as much to run . What it means for SMBs / MSPs: Open-weight models now offer credible frontier-quality alternatives at a fraction of frontier pricing—consider evaluating for cost-sensitive deployments where you can run locally or manage your own tuning.
- Gemini 3.5 Pro General Availability Today (Leaked Launch Plans)
Gemini 3.5 Pro launches today with a 2-million-token context window, Deep Think reasoning on the $250 per month Ultra tier, and API pricing around $1.25 input and $10 output per million tokens .
At $1.25 input with 2 million tokens of context, entire categories of RAG architecture get simpler because you can just put the corpus in the prompt . What it means for SMBs / MSPs: This pricing and context window make RAG workflows simpler—you may not need complex pipeline design if your document corpus fits in context. Worth benchmarking against Claude Sonnet 5 for document-heavy workloads.
- Moonshot AI K3 and Kimi Integration in GitHub Copilot (LLM News)
Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter open MoE model built on Kimi Delta Attention, activating 16 of 896 experts .
Kimi K2.7 Code is now available directly inside GitHub Copilot, giving cost-aware teams a cheaper route with usage-based billing . What it means for SMBs / MSPs: Open-weight coding models in GitHub Copilot reduce vendor lock-in on Copilot Chat—test Kimi K2.7 for teams sensitive to per-token costs on coding tasks.
- Claude Sonnet 5 Intro Pricing Extended Through August 31 (AIapps)
Claude Sonnet 5 from Anthropic brings stronger long-run coding, tool use, and debugging at a lower price .
It became the default model for all Claude plans on June 30, 2026 , with
intro pricing at a lower tier through Aug. 31, 2026, making it a solid pick for teams trying out autonomous coding and workflow automation without spending top-tier money . What it means for SMBs / MSPs: Lock in Sonnet 5 pilots by August 31 if you want to test agentic coding workflows at promo rates—after that, pricing normalizes.
- Microsoft 365 Business with Copilot Now Permanent SKU (Microsoft Partner Center)
As of July 1, 2026, Microsoft 365 Business with Copilot SKUs are now generally available, turning a successful promotional motion into a permanent, predictable offer .
Anthropic's Claude is now available in Copilot Chat, giving users more flexibility in how they work . What it means for SMBs / MSPs: You can now sell Copilot as a permanent fixture in M365 Business bundles rather than a time-limited promo—Claude option in Copilot Chat gives your customers model choice without leaving the Microsoft ecosystem.
Claude Sonnet 5 via Anthropic's API or Claude.ai
Claude Sonnet 5 is the highest-value model for agentic workflow testing this week.
It's the most practical model to test first if you want something you can use right away across a lot of tasks, as it became the default model for all Claude plans on June 30, 2026 .
Who it's for: SMB owners, AI champions, and MSPs piloting autonomous coding and multi-step workflows.
First test timeline: 2–3 hours to set up a proof-of-concept (e.g., automated document processing, lead qualification, or code review).
Insurance tech firm Pace used it on live systems for multi-step insurance work, including intake and claims setup .
First step: Pick one high-volume, repetitive task your team handles manually. Write a prompt that asks Claude Sonnet 5 to break it into steps, execute them, and return a structured output. Measure execution time, error rate, and whether a human needs to spot-check each run. Compare against your current process cost.
List the long documents your team reads by hand each month, like contracts or manuals. Gemini 3.5 Pro, generally available today, can take a very large document set in a single prompt. Sonnet 5's introductory pricing now runs through August 31, so trying one of those documents is cheaper this summer. Check the tool's data terms before you upload anything sensitive.
*For MSPs:* Use the Sonnet 5 introductory pricing window to prove workflow value with clients before full pricing starts.
Don't bet your customer's workflow on "frontier model of the moment" benchmarks.
For most businesses the winning move is not chasing whichever model leads the benchmarks this month, it is building on a flexible setup you can swap newer models into as they improve . This week alone saw Claude Sonnet 5, GPT-5.6, Grok 4.5, Gemini 3.5 Pro, and Kimi K3 all land within days—if you commit architecturally to one, you'll rebuild in 90 days anyway.
Open-weight models are moving from research artifacts to production credentials.
Researchers from Thinking Machines and Bridgewater Associates took an existing open-source model and trained it further on Bridgewater's own financial expertise, scoring 84.7% on financial reasoning tests, beating top proprietary models, while costing roughly a fourteenth as much to run .
Mid-tier models like Claude Sonnet 5 and GPT-5.6 Sol are getting close to flagship-level performance at lower prices, and open-weight options like Kimi K2.7 Code, now available directly inside GitHub Copilot, give cost-aware teams a cheaper route with usage-based billing . The practical implication:
Frontier quality is now a moving target shared by several providers, with open source close behind. Stop trying to pick a permanent winner. Build a model-agnostic foundation, test new releases against your own tasks, and route each job to the model that fits. Do that and a market that changes every three days becomes a steady stream of upgrades working in your favour .