AI News: Claude Sonnet 5.5, GPT-6.1 and Open Models

Claude Sonnet 5.5 and GPT-6.1 Sol now share the same $2 input and $10 output price. Procurement now turns on tooling fit, open-weight deployment and operational control as near-frontier intelligence gets cheaper.

Price stopped separating the main coding models

Anthropic released Claude Sonnet 5.5 on September 28, followed by OpenAI's GPT-6.1 Sol at DevDay on September 29. Both cost the same at standard rates, but their API behavior differs enough to make benchmark-based switching careless.

Anthropic reports that Sonnet improved from 10.3% to 70.6% on Terminal-Bench 4.0, a benchmark for agents operating through terminal environments. Teams running with thinking disabled must migrate to the new between_tools setting because existing configurations won't behave identically.

GPT-6.1 Sol introduces different constraints:

• Prompts above 272,000 input tokens cost twice the normal input rate.
• The same oversized request pays 1.5 times the normal output rate.
• Supported reasoning.effort values run from low to max.
• Chat Completions supports Sol only without tool calling.
• Agent systems that call tools need the Responses API.

OpenAI says Sol approaches GPT-6 Astra at one-fifth of Astra's standard token price, according to the DevDay launch coverage. Most engineering teams should compare task completion rate, retry frequency and tool-call reliability alongside the published base token price.

Haiku 5.5 makes subagent routing economically obvious

Anthropic released Claude Haiku 5.5 on October 7 and positioned it below Sonnet and Opus for complex agentic coding. Its target workloads are compaction, summarization and focused subagent tasks, according to the Haiku 5.5 announcement.

The price gap is too large to ignore:

• Input: $0.10 per million tokens
• Output: $0.50 per million tokens
• Standard-price prompt limit: 100,000 tokens
• Recommended work: summaries, context compression and narrow delegated tasks

Use Haiku for bounded, high-volume tasks with validation. Escalate failed or uncertain outputs to Sonnet or Opus for ambiguous planning and difficult code changes. Sending every intermediate step to Sonnet 5.5 costs 20 times more for input and output without enough improvement in the final answer to justify it.

Mistral Large 4 challenges hosted models before its license is known

Mistral placed Large 4 into public preview on October 6 through the Mistral Studio API. The company describes it as a mixture-of-experts model with roughly one trillion total parameters and 52 billion active for each token, according to the official release.

Inference cost depends on active parameters, not total parameter count. Running only part of the model for each token can deliver large-model capacity without activating the full network on every request.

The API pricing is aggressive, but teams planning private deployment should wait for the actual license, serving requirements and memory footprint before committing architecture or budget:

• Input: $1.36 per million tokens
• Output: $4.18 per million tokens
• Current status: public API preview
• Weight release target: end of October
• License: not yet announced

Beam and Kolibri show two very different open-weight strategies

Reflection announced Beam on October 5, while Aleph Alpha released Kolibri on October 3. Both use sparse mixture-of-experts architectures. Kolibri is available to download now; Beam remains waitlisted.

Reflection reports an 80.9 score on SWE-Bench Verified for Beam. The model remains text-only and waitlisted until the promised Apache 2.0 release arrives later in October.

Kolibri is the less spectacular but more actionable release. It supports English and German, offers context up to one million tokens and is already published under Apache 2.0, though serving requires Aleph Alpha's inference plugin for vLLM.

Kolibri deployments beyond 262,144 tokens need extra serving flags. Teams should test memory use and latency at their real context lengths instead of treating the maximum specification as the normal operating point.

Cloudflare Clef argues that many AI calls should not return prose

Cloudflare released Clef and Clef-flash on October 1 as Apache 2.0 decision models. They return bounded, typed outputs with probabilities, as detailed in the Clef announcement.

The interface suits routing, policy checks, classification and workflow gates:

• Models are available through Workers AI as @cf/cloudflare/clef.
• Clef-flash reports 38.8 ms median latency.
• Outputs use defined types.
• Responses include probabilities that applications can compare against thresholds.
• Reinforcement learning fine-tuning starts as an engineer-run service, with self-service access planned.

This is the most important architectural release in the group. Small decision models are often the better production component when output must drive code. A support workflow can route high-confidence billing requests automatically while sending low-confidence cases to a person, without asking a general chat model to produce JSON and hoping it follows the schema.

Worth watching

• Whether Mistral publishes Large 4 weights by the end of October, and which license accompanies them.
• Whether Reflection ships Beam's Apache 2.0 weights on schedule and provides reproducible evaluation details.
• How OpenAI customers manage Sol's pricing boundary once production agents begin filling larger contexts.
• When Cloudflare opens self-service reinforcement learning fine-tuning for Clef.
• Whether model routers begin treating Haiku-class models as the default worker tier.

What To Do Now

1. Run the same 20 repository tasks through Claude Sonnet 5.5 and GPT-6.1 Sol, recording successful completion, tool failures, retries and total token cost.
2. Audit one week of GPT request logs and count how many prompts exceed 272,000 input tokens before adopting Sol for long-context workflows.
3. Route 100 summarization or context-compaction jobs to Haiku 5.5, then compare validation failures and cost against your current model.
4. Test Kolibri in an isolated environment with its required vLLM plugin before proposing it for production or regulated workloads.
5. Prototype one Clef decision endpoint with a fixed schema, then log confidence scores and p95 latency for at least 1,000 requests.

Useful Resources

• Claude Sonnet 5.5 - Release details, pricing and agent benchmark results.
• GPT-6.1 Sol API documentation - Context, pricing thresholds and API compatibility.
• Claude Haiku 5.5 - Pricing and recommended workload categories.
• Mistral Large 4 - API preview details and planned weight release.
• Cloudflare Clef - Typed decision outputs, latency and Workers AI availability.

Looking Ahead

Equal headline prices expose the architectural differences underneath them, from tool APIs and context surcharges to deployment licenses and typed outputs.

The winning stack won't send everything to one flagship model. It will route narrow work to cheap subagents, reserve expensive reasoning for hard decisions, and use deterministic decision models where software needs a bounded answer. That is the same control problem behind the recent agent control fight, and these releases make it an immediate engineering priority.