Gemini API Prices Double December 1; Fractile Raises $720M for Inference Chips

Promotional pricing ends for major LLM APIs

Google will double Gemini 4 Argon prices on December 1, 2026, moving from $2 per million input tokens and $10 per million output tokens to $4 and $20 respectively. Gemini 3.6 Flash and 3.8 Flash rates—currently $0.75 input and $3.75 output—are scheduled to double again on January 1, 2027. OpenAI's GPT-6.1 Sol and xAI's Grok 4.7 maintain $2 input pricing but vary significantly on output: $10 and $6 per million tokens respectively, with both adding context-length surcharges above thresholds.

The practical impact is sharp. A workload sending 1,000 requests with 3,000 input and 500 output tokens per request costs between $0.55 and $55 depending on model selection. For a procurement team that budgeted based on current Gemini promotional rates, December's price change doubles run-rate costs unless the application shifts to smaller models, adds routing, or migrates to self-hosted infrastructure.

Buyers should reforecast 2027 API spend using post-promotion pricing and negotiate volume tiers, price-protection clauses, and model-substitution rights. Calculate cost per completed workflow or accepted answer rather than raw token consumption—output quality and task completion matter more than tokens processed.

Fractile raises $720 million to challenge NVIDIA in inference

British AI-chip company Fractile closed a $720 million Series C on November 12 to develop purpose-built LLM inference accelerators and a machine-learning compiler. The company targets the same inference workloads currently dominated by NVIDIA GPUs, positioning against both GPU incumbents and emerging competitors including AMD, Groq, Cerebras, and custom hyperscaler silicon.

The financing is material because it strengthens the case for evaluating heterogeneous inference fleets. If Fractile or similar vendors deliver competitive tokens-per-dollar and latency under production concurrency, enterprises can reduce dependency on GPU supply chains and potentially lower inference costs for stable, high-volume workloads.

However, funding does not equal deployment readiness. Buyers evaluating Fractile or other ASIC-based inference platforms should verify chip availability, supported model architectures, framework compatibility, measured throughput under concurrent load, compiler maturity, and migration costs if the platform lags CUDA-based tooling in future model releases. Purpose-built accelerators trade flexibility for efficiency—an acceptable trade for predictable production workloads, a risk for exploratory or rapidly evolving applications.

Rackspace-Together AI infrastructure plan lacks committed GPU financing

Rackspace and Together AI announced a strategic agreement to support enterprise AI inference using NVIDIA GB300 NVL72 systems, with deployment planned for 2027. Rackspace separately secured a $66 million private placement led by Voss Capital on November 6. However, Rackspace disclosed it has no definitive GPU-financing commitment and still needs separate capital to acquire the GB300 hardware.

Enterprises seeking dedicated or managed inference capacity should treat this as a potential 2027 supply option, not immediately available infrastructure. The absence of committed GPU financing increases execution risk. Buyers evaluating proposals from Rackspace or similar providers should demand contractual guarantees around delivery dates, capacity availability, utilization commitments, and pricing—especially for hardware that is not yet funded.

Compare the total cost of dedicated GB300 capacity—including financing, power, networking, software licensing, and idle capacity—against API-based inference from hyperscalers and reserved instance pricing from AWS, Azure, and Google Cloud. Dedicated infrastructure makes sense when workload volume justifies fixed costs and when latency or data-residency requirements rule out shared APIs. It does not make sense when capacity is speculative.

EnterpriseBench tests agent planning, not just accuracy

Researchers published EnterpriseBench on November 12, a benchmark designed to evaluate LLM agents on strategic reasoning and dynamic decision-making rather than static question answering. The benchmark tests whether an agent can plan, act, use tools, recover from errors, and make decisions in business contexts—capabilities procurement teams need before granting agents production permissions.

The benchmark complements existing software-engineering and coding evaluations but competes with vendor-controlled demonstrations that may favor particular models or workflows. Buyers should require vendors to disclose performance on independent, enterprise-relevant tasks before approving autonomous workflows. Testing planning reliability, tool use, and decision quality adds evaluation cost and time but reduces the risk of deploying agents that fail under ambiguity or multi-step reasoning.

EnterpriseBench is more useful as an evaluation-design signal than as evidence that one model currently leads. No universally accepted leaderboard exists yet. However, the publication indicates that enterprise agent evaluation is moving beyond accuracy metrics toward operational readiness.

What to do now

Reforecast 2027 API budgets using Gemini's December 1 pricing and January 1 Flash increases. Negotiate price protections, volume tiers, and model-substitution rights into enterprise contracts. Evaluate routing systems that send simple requests to cheaper models and reserve premium models for complex tasks.

Benchmark inference on at least one GPU option and one specialized accelerator before committing to long-lived infrastructure. Verify that announced capacity—such as Rackspace's GB300 plan—has committed financing and delivery dates.

Require vendors to disclose agent performance on independent benchmarks such as EnterpriseBench before approving autonomous workflows. Test planning reliability, tool use, and error recovery under production-like conditions, not just demonstration scenarios.