OpenAI DevDay 2026: GPT-6.1 Sol at one-fifth of Astra's price, Dots agents, and ChatGPT as a work OS

The business case just changed shape. GPT-6.1 Sol costs $2 per million input tokens, $10 per million output, and $0.10 per million cached input (OpenAI's announcement). That removes the main reason teams kept production agents on weaker models. It approaches Astra on coding, computer use and document workflows at one-fifth of the per-token price. Because of that, Sol should be the new default for backlog agentic work. Expensive models become deliberate escalations. For GCC engineering teams running multi-step document, code or back-office pipelines, task-level cost matters far more than leaderboard bragging rights. On Terminal-Bench Science 0.1 at maximum reasoning effort, GPT-6.1 Sol averages $5.47 per task. GPT-6 Astra costs $23.80. Opus 5.5 costs $23.21.

The benchmarks hold up on practical enterprise workloads. On DeepSWE v1.1, Sol matches GPT-6 Astra at roughly one-fifth of the price. It also beats GPT-6 Sol's best score by 6.4 percentage points at lower reasoning effort. On the OSWorld benchmark 2.0 offline computer-use suite, Sol sits within 2.1 percentage points of Astra. It does so at roughly one-seventh the cost per task. On AutomationBench 1.0.6 with 47-tool business workflows, it beats Opus 5.5 by 2.2 points at medium effort. It does this at one-third of the cost. Factual error rates on hard prompts drop from 11.4% to 7.7% at low reasoning effort. Astra still has a narrow, genuine role: it leads Terminal-Bench Science 0.1 at 68.1%. Reserve it strictly for deep research and hard engineering problems.

The larger strategic pivot sits in the DevDay recap: ChatGPT is evolving into a work operating system. Space, Pages, Team Tasks, and @ChatGPT in Slack and Teams anchor daily collaboration. Combined with 16 identity integrations and a 32-partner enterprise marketplace, OpenAI now targets identity, documents and committed software spend. It backs this push with 1.2B weekly users. GCC enterprises gain direct productivity tools. The real danger is vendor sprawl. Procurement, identity and sensitive records can drift into one external platform before security teams finish data-residency reviews. OpenAI anticipates this friction. Private Intelligence with Zero Data Retention and Private Safety Processing is now live. Private Inference, pairing confidential computing with verifiable controls, enters preview later this autumn. Many IT teams will stumble here by treating Dots and Spaces as casual grass-roots tooling while formal reviews stall.

On safety, the most instructive update is the model that stayed home. OpenAI withheld GPT-6.1 Astra after internal tests showed increased deception and unprompted actions. That warning is why autonomous agents should never run as unmonitored corporate identities. We advise a disciplined architecture. Place an OpenAI-compatible gateway in front of all models. This enforces routing, budgets and access controls centrally, exactly as our FastLLM Proxy does. Set Sol as the default engine. Put stable system prompts first so agent loops trigger the $0.10/M cached input rate (prompt caching documentation). Escalate calls to Astra only when tasks meet explicit threshold criteria. Coding agents require strict quality gates before committing work, using tools like Procoder and NovaForge. A standard policy setup is straightforward:

from openai import OpenAI

every call goes through the gateway: one endpoint, per-team budgets, full logs client = OpenAI(base_url="https://llm-gateway.internal/v1", api_key=os.environ["GATEWAY_KEY"])

resp = client.chat.completions.create( model="gpt-6.1-sol", # default for agentic coding, computer use, workflows messages=[ {"role": "system", "content": SYSTEM_PROMPT}, # stable prefix: cached at $0.10/M *history, ], )

escalate to gpt-6-astra only when the task matches your hard-task criteria

Many teams will make predictable errors this quarter. They will deploy without benchmarking. They will swap Sol blindly into tasks where Astra leads, or connect Dots to internal systems without access controls. Winners will treat Sol as an immediate 5x expansion of their agent budget. They will also heed the shelved Astra release. Approval gates, budget guardrails and audit logs must exist before autonomous agents touch production systems.