OpenAI, Google, Anthropic, and Meta: The Model Releases Worth Watching This Week

What happened this week

Anthropic: Claude Opus 5.5

Anthropic launched Claude Opus 5.5 on September 22, the first model in its new Claude 5.5 family. The headline numbers: 40% lower cost than Opus 5 on typical workloads, output speed more than 30% faster, and token pricing down to $4 input / $20 output per million tokens—20% below Opus 5.

The benchmark results were immediate. Opus 5.5 scored 58 on the Artificial Analysis Intelligence Index, the highest recorded score to date on that platform. On BenchLM's reasoning shortlist, it sits at an estimated 82.4 reasoning score, behind only GPT-6 Astra (89.5) but ahead of Claude Fable 5.1 (81.4).

The practical takeaway: Opus 5.5 performs at the level of Fable 5.1—Anthropic's previous top-tier model—at roughly 60% of the cost. One early tester completed a 680,000-line code migration in less than a day, work that would have taken an engineering team weeks.

Skip it if: You need the absolute highest reasoning ceiling. GPT-6 Astra still leads that category. But if cost efficiency at near-frontier quality is your priority, Opus 5.5 is the strongest option available right now.

OpenAI: GPT-6 Sol and GPT-6 Luna

OpenAI expanded its GPT-6 lineup on September 23 with two models priced at half the promotional rates of their predecessors, while inheriting some capabilities from the flagship GPT-6 Astra. The launch followed GPT-6 Astra's release earlier in September.

GPT-6 Sol and Luna were trained using methods similar to Astra, with improvements in reasoning, factual reliability, coding, computer use, and alignment. OpenAI also said the new models are significantly less likely to attempt to bypass restrictions than previous generations.

The pricing move is aggressive and permanent—not a limited-time discount. API input prices dropped to $2 per million tokens for these models. On BenchLM's reasoning index, GPT-6 Sol scores an estimated 79.6, behind Astra (89.5) and Opus 5.5 (82.4), but at only $4 for the same workload.

The practical takeaway: If Astra was too expensive for your workflow, Sol and Luna are the answer. OpenAI CEO Sam Altman framed the price cut as a step toward broader AI adoption.

The tradeoff: Astra remains the most capable model for demanding projects. Sol and Luna are optimized for professional work, coding, automation, and computer-use tasks—not for pushing the reasoning frontier.

Google: Gemini 3.8 Live and Live Avatar

Google took a different path. Rather than chasing the intelligence leaderboard, Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking in mid-September, focused on real-time voice dialogue and multi-step reasoning through voice.

Then on September 24, Google layered on Live Avatar—a system that combines near real-time video generation with speech, giving the voice model a visual presence. The avatar supports lip-sync and facial expressions across 97 languages.

The benchmark results support the investment. Gemini 3.8 Live Extended Thinking ranked first on Artificial Analysis's Speech-to-Speech Quality Index with 82.6 points, while Gemini 3.8 Live ranked second in the Speech Agent Arena.

The practical takeaway: Google is not competing on the same axis as OpenAI and Anthropic. If you need a real-time voice assistant for enterprise deployment—customer support, accessible interfaces, hands-free workflows—Gemini 3.8 Live is currently the strongest option.

Skip it if: You need text-based reasoning or coding assistance. This release is not designed for those tasks.

Meta: Muse Spark 1.3 and Muse Glimmer 30B

Meta's release was quieter but worth noting. The company expanded its Muse family with Muse Spark 1.3, which features a 1-million-token context window and is priced at $1.25 / $4.25 per million tokens. Meta also released Muse Glimmer 30B, a smaller model that currently leads Meta's public score ranking at 45.08.

These are not frontier models. But Meta's strategy is different: open-weight availability and integration into its ecosystem—Facebook, Instagram, WhatsApp, Messenger, and smart glasses.

The practical takeaway: If you are building on open-weight models or developing within Meta's ecosystem, Muse Spark 1.3 is the model to evaluate. If you need top-tier reasoning, look elsewhere.

The bigger picture: cost is the new battleground

The most significant pattern this week is not any single model. It is that every major provider moved toward cheaper, faster, more deployable models at roughly the same time.

Anthropic cut costs by 40%. OpenAI cut costs by 50%. Google focused on cost-efficient voice agents at scale. Meta priced aggressively for ecosystem adoption. As one industry analysis put it, the competition has shifted from "whose model is smarter" to who can deploy near-frontier capability at lower cost and higher speed into more real tasks.

For model selection, this means the default question is changing. It is no longer "which model is best?" It is "which model is best for this task at this budget with acceptable speed?" The answer this week leans heavily toward the cheaper tier of models for most professional work.

FAQ

Which AI model is best right now for reasoning?

GPT-6 Astra leads BenchLM's reasoning estimate at 89.5, followed by Claude Opus 5.5 at 82.4 and Claude Fable 5.1 at 81.4. But "best" depends on your budget. GPT-6 Astra costs $20 for the stated workload; Opus 5.5 costs $8 for the same workload at a lower score.

Is GPT-6 Sol better than Claude Opus 5.5?

They serve different needs. Opus 5.5 scores higher on reasoning estimates (82.4 vs 79.6 for Sol) and is optimized for complex coding and agentic tasks. GPT-6 Sol is optimized for cost-efficient professional work and automation at a lower price point. If your task is coding-heavy, Opus 5.5 has the edge. If you need volume at low cost, Sol is the better fit.

What is Gemini 3.8 Live best for?

Real-time voice assistants. It ranked first on Artificial Analysis's Speech-to-Speech Quality Index at 82.6 points and supports 97 languages with near real-time visual input. If you are building a voice-first product or need hands-free AI interaction, this is the current leader.

Are these price cuts permanent?

OpenAI confirmed that GPT-6 Sol and Luna pricing is long-term, not a limited-time promotion. Anthropic's Opus 5.5 pricing reflects a structural reduction in compute requirements—it needs less compute to serve than Opus 5, so the lower price is sustainable.

Which model should a small business with no technical team choose?

For general professional work—writing, research, document analysis—GPT-6 Sol offers the best balance of capability and cost. For coding or complex technical tasks, Claude Opus 5.5 is worth the premium. If your workflow is voice-based, Gemini 3.8 Live is the clear choice.

What should I watch next week?

Claude Sonnet 5.5 and Haiku 5.5 are expected to follow Opus 5.5 in the coming weeks, which will likely bring the 5.5 family's cost and performance improvements to lower price tiers. Watch for pricing updates across the mid-tier model segment.