GPT-6.1 Sol and Sonnet 5.5 arrived one day apart, each promising near-flagship performance for a fraction of the price. The real news isn't either model, it's that the industry finally admitted its top-tier pricing was never sustainable.
One day apart, OpenAI and Anthropic each cut the price of near-flagship AI by roughly 80%. On September 29, at DevDay 2026 in San Francisco, OpenAI released GPT-6.1 Sol, a model it says delivers "near-Astra intelligence" at a fifth of GPT-6 Astra's standard token prices. It's priced at $2 per million input tokens and $10 per million output tokens, with cached input running $0.10 per million, which OpenAI describes as 95% below its standard input rate. The day before, on September 28, Anthropic shipped Claude Sonnet 5.5 at the exact same headline price, $2 and $10 per million tokens, unchanged from Sonnet 5, but says it now runs more than 30% faster and costs up to 30% less per completed task because it burns fewer tokens and makes fewer tool calls to finish the same work.
Two labs, one week, the identical pitch: don't pay flagship prices to get flagship-adjacent results. That's not a coincidence. It's what happens when GPU capacity gets cheaper faster than demand does, and every lab is racing to own the price point where actual production traffic lives.
The numbers back up the pitch, mostly. On DeepSWE v1.1, a coding benchmark, GPT-6.1 Sol scores 75% versus Astra's roughly 74.8%, essentially a tie, while OpenAI says it cuts the cost per task by about 80%. On OSWorld 2.0, a computer-use benchmark, Sol comes within 2.1 percentage points of Astra's 73.5% flagship score while costing around $1.30 per completed task against Astra's $9.30, per benchmark data reported by Vellum AI. That's the case for Sol in a sentence: close enough on the tasks that pay the bills, dramatically cheaper.
But the gap widens on harder problems. On Terminal-Bench Science, Sol scores 57% at $5.47 per task, while Astra hits 68.1% for $23.80. Eleven points is not nothing if you're running scientific reasoning workloads at scale. Sol also arrived just one week after OpenAI's previous cheap-tier model, GPT-6 Sol, itself scored 68.8% on DeepSWE, so most of Sol's coding gains actually came from that prior release, not this one. If your work leans toward frontier reasoning rather than coding and agentic tasks, Astra still earns its higher price. If you're shipping code or running browser agents, Sol is the more rational default now.
Anthropic wants a $2 trillion price tag on a company that lost $42 billion last year Anthropic's IPO prospectus discloses a $42 billion net loss on $4.6 billion in 2025 revenue, but roughly $34 billion of that is a non-cash accounting charge tied to convertible instruments. The company is still targeting a $2 trillion valuation while committing to $518 billion in future cloud spending. - Anthropic IPO valuation two trillion dollars loss - AI company IPO prospectus billion dollar losses revenue
Availability is broad and immediate. Sol is live for Plus, Pro, Business, Enterprise, and Edu users in ChatGPT and Codex, and developers can call it through the API as gpt-6.1-sol starting September 29. An Ultrafast variant, running up to eight times faster in Codex, is coming in the following days.
Sonnet 5.5's bet: speed over discount #
Anthropic took a different route to the same destination. Rather than cutting the sticker price, Sonnet 5.5 keeps Sonnet 5's rate card and instead squeezes more work out of the same dollar. On Terminal-Bench 4.0, an agentic coding benchmark, Anthropic reports Sonnet 5.5 scoring 70.6%, ahead of its own Opus 5.5 flagship at 66.4% and miles past the previous Sonnet 5's 10.3%. On GDPval-AA v2.1, a knowledge-work evaluation, Sonnet 5.5 scored 1,844, just two points behind Opus 5.5's 1,846 and about 400 points above Sonnet 5.\n\nThat's a striking claim on its face: a mid-tier model beating its own flagship on a coding benchmark. It says less about Sonnet 5.5 being smarter than Opus and more about what Anthropic optimized for, fewer wasted tool calls and tighter reasoning traces on tasks where brute intelligence matters less than execution discipline.
Frankly, this is the more interesting release of the two, because Anthropic isn't just discounting, it's arguing that price and quality decoupled from raw model size a while ago. Sonnet 5.5 shipped September 28, priced identically to its predecessor, betting that efficiency sells better than a lower number on the invoice.
Then there's the open-weight pressure underneath both of them. Alibaba's Qwen3.8-27B, available through OpenRouter, runs as low as $0.0427 per million input tokens and $4.40 per million output tokens, and Alibaba's own API prices it at $0.50 and $3.00. It scores 34 on the Artificial Analysis Intelligence Index, ranking 32nd on that firm's coding index, well below both Sol and Sonnet 5.5 on raw capability. But for founders building high-volume, low-complexity pipelines, classification, extraction, simple agent loops, Qwen3.8-27B is already cheap enough that the frontier labs' new
Also read: Anthropic warns a free Chinese AI model can already build working hacks • OpenAI launches Dots to rival Meta's Muse and it stumbles on stage • General Compute adds Cerebras chips to its Nvidia fleet to chase faster AI coding agents
This article is posted in AI News, check it out for more related stories.
OpenAI and Anthropic Are Quietly Probing Tens of Thousands of AI Security Incidents Axios reports that OpenAI and Anthropic are investigating tens of thousands of security incidents involving their AI models and agents, most never disclosed publicly. The finding follows a month of individual failures, from a DNS-based sandbox escape at OpenAI to a nine-zero-day breach of Hugging Face, and raises hard questions about whether any... - how AI models escape sandbox security measures - anthropic and openai security incident investigation details
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.