OpenAI and Anthropic Extend 2026 Model Race With GPT-6 Sol, GPT-6 Luna and Claude Fable 5.1

OpenAI and Anthropic have turned 2026 into a rapid-release contest, moving from agentic coding upgrades to cheaper frontier models, larger context windows and more explicit safety tiers. OpenAI's latest step came on September 22 with GPT-6 Sol and GPT-6 Luna, while Anthropic's recent cadence has pushed Claude through Sonnet, Opus, Fable and Mythos releases. OpenAI introduced GPT-6 Sol and GPT-6 Luna as lower-cost GPT-6 variants following GPT-6 Astra earlier in September. The new models are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, with Free and Go users able to access GPT-6 Luna in the desktop app; API developers can call gpt-6-sol and gpt-6-luna. OpenAI priced GPT-6 Sol at $2 per 1 million input tokens and $10 per 1 million output tokens, and GPT-6 Luna at $0.10 input and $0.50 output, saying those rates are 50% below GPT-5.6 promotional pricing. The company also said prompt caching for GPT-6 offers 90% discounts on cached input-token reads. OpenAI framed Sol and Luna as cost-efficiency plays rather than replacements for Astra. Its published evaluations say GPT-6 Sol at xhigh effort scored 33.2% on AutomationBench at $0.27 per task, compared with 26.9% for Claude Opus 5 at 11.1 times the cost and 31.4% for Claude Fable 5.1 with Opus 5 fallback at more than 8.9 times the cost. On DeepSWE v1.1, OpenAI reported GPT-6 Sol at 68.8%, within 1.1 percentage points of Claude Fable 5's 69.9% score, at about 80% lower cost per task; GPT-6 Luna at max effort scored 66.6%. The September launch caps a long OpenAI sequence. The company released GPT-5.3-Codex on February 5 for paid ChatGPT plans across the Codex app, CLI, IDE extension and web, and said it was running 25% faster for Codex users. It followed on March 5 with GPT-5.4 and GPT-5.4 Pro, positioning GPT-5.4 as a professional-work model with native computer-use capabilities, up to a 1 million-token context window in Codex, 83.0% on GDPval and 75.0% on OSWorld-Verified, above the 72.4% human baseline reported there. On March 17, OpenAI added GPT-5.4 mini and GPT-5.4 nano for lower-latency coding, multimodal and subagent work; GPT-5.4 mini was priced at $0.75 input and $4.50 output per 1 million tokens, while GPT-5.4 nano was API-only at $0.20 input and $1.25 output. OpenAI introduced GPT-5.5 on April 23, later adding GPT-5.5 and GPT-5.5 Pro to the API, with GPT-5.5 priced at $5 input and $30 output per 1 million tokens and a 1 million-token context window. On July 9 it released the GPT-5.6 family: Sol, Terra and Luna. GPT-5.6 Sol posted an Artificial Analysis Coding Agent Index score of 80, 62.6% on OSWorld 2.0, 96.7% on Capture-the-Flag challenges and 94.6% on GPQA Diamond in OpenAI's tables. Anthropic's 2026 sequence has emphasized agentic behavior, enterprise workflows and graduated safeguards. Claude Sonnet 4.6, released February 17, became the default model for Claude Free and Pro plans, added a 1 million-token context window in beta and kept Sonnet 4.5 pricing at $3 input and $15 output per 1 million tokens. Anthropic said early Claude Code users preferred Sonnet 4.6 to Sonnet 4.5 roughly 70% of the time and to Opus 4.5 59% of the time. On May 28, Anthropic introduced Claude Opus 4.8, citing improvements over Opus 4.7, a 2.5 times faster fast mode and lower fast-mode pricing. The company then launched Claude Fable 5 and Claude Mythos 5 on June 9 at $10 input and $50 output per 1 million tokens. Fable 5 was generally available with classifiers that routed some cybersecurity, biology, chemistry and distillation requests to Claude Opus 4.8; Anthropic said more than 95% of Fable sessions involved no fallback. Mythos 5 used the same underlying model with fewer safeguards for selected Project Glasswing partners. That safety split quickly became the year's clearest limitation story. Anthropic suspended access to Fable 5 and Mythos 5 on June 12 after a US government directive tied to concerns about a jailbreak, then restored Fable 5 globally after the controls were lifted. Anthropic said its testing found the reported behavior did not expose unique Mythos-level cyber capabilities and that other models could identify the same vulnerabilities. It later added Claude Sonnet 5 on June 30 at $2 input and $10 output per 1 million tokens, making it available across all plans and describing it as close to Opus 4.8 performance at lower prices. Claude Opus 5 followed on July 24 as the default model on Claude Max and the strongest model on Claude Pro, with Anthropic saying it came close to Claude Fable 5 intelligence at half the price. Anthropic's September update, Claude Fable 5.1 and Claude Mythos 5.1, shows how customer reception is shaping the race. Anthropic said Fable 5.1 addresses feedback on price, data retention and safeguards: typical token-billed workloads cost an estimated 25% less than Fable 5, highly agentic work can save up to about 45%, and cybersecurity false positives are reduced by 60%. It also said Fable 5.1 scored 52.6% on Terminal-Bench-Science 0.1 versus 24.7% for Fable 5, and 55.8% on Terminal-Bench 4.0, while Mythos 5.1 reached 60.9% on the same coding benchmark. The practical reception is mixed in the way frontier AI adoption usually is: developers and enterprise testers are rewarding models that can stay on task, use tools, finish code changes, build documents and run longer workflows with fewer handoffs, while buyers are pressing for lower per-task cost, better cache economics, clearer data-retention terms and fewer benign requests caught by safeguards. The outlook is that model competition is no longer just about a single top benchmark. The companies are increasingly competing on the full operating envelope: agent autonomy, context length, coding reliability, computer use, prompt caching, token price, fallback behavior and whether advanced cyber or biology capabilities are opened broadly, gated through verification programs or kept to trusted-access deployments.