Frontier Models

Overview

Frontier AI model development is no longer dominated by a single architectural approach or a single lab. As of early 2026, major labs are experimenting with inference-time scaling — using multiple agents at generation time rather than simply training bigger models — and domain-specific specialization. Meanwhile, open-weight models running on consumer hardware are closing the capability gap faster than expected.

The November 2025 Inflection Point

November 2025 was a critical month for frontier models. The "best" model — judged largely on practitioner vibes — changed hands five times: Claude Sonnet 4.5 (September) was overtaken by GPT-5.1, then Gemini 3, then GPT-5.1 Codex Max, before Anthropic reclaimed the lead with Claude Opus 4.5 source(https://simonwillison.net/2026/May/19/5-minute-llms/). Opus 4.5 held the consensus crown for roughly the next two months.

More consequentially, November was when coding agents crossed a quality threshold. OpenAI and Anthropic had spent most of 2025 applying Reinforcement Learning from Verifiable Rewards (RLVR) to improve code generation, particularly when paired with their Codex and Claude Code agent harnesses. The result: coding agents went from "often-work" to "mostly-work," reaching the point where practitioners could use them as a daily driver without spending most of their time fixing mistakes source(https://simonwillison.net/2026/May/19/5-minute-llms/). See Agentic Engineering and Agent Proficiency for more on this shift.

Meta's Pivot: Open to Closed (Muse Spark)

Meta released Muse Spark in April 2026, its first AI model in over a year and the first from its nine-month-old Superintelligence Labs. It represents a significant strategic shift: after years of championing open-weights models (Llama), Meta released a proprietary closed model.

Background: Meta reorganized its AI labs after critics alleged Llama 4 training data was contaminated with benchmark answers. In June 2025, Meta spent $14.3B for a 49% stake in Scale AI, brought in cofounder Alexandr Wang as chief AI officer, and launched a major hiring spree. The proprietary release has raised concerns among developers who built projects on Llama.

Architecture highlights:

• Natively multimodal: text, image, speech input (up to 262K tokens); text output
• Three reasoning modes: instant, thinking, contemplating
• Multi-agent contemplating mode: launches multiple agents that propose solutions, refine them, and aggregate results in parallel
• Post-training used RL penalizing excessive reasoning tokens ("thought compression"): model first improved by reasoning longer, then learned to compress, then extended again
• Reworked pretraining approach, model architecture, optimization, and data curation — claims to match Llama 4 Maverick's capabilities with >10x less training compute
• Health specialization: >1,000 physicians helped curate training data for health responses
• Parameter count, architecture, and training details undisclosed

Availability: Free via meta.ai and Meta AI app; API preview for selected partners; coming to WhatsApp, Instagram, Facebook, Messenger, Ray-Ban glasses.

Benchmark Snapshot (Artificial Analysis, April 2026)

Muse Spark's notable strengths: multimodal (CharXiv Reasoning: 86.4%, leading GPT-5.4's 82.8%) and health (HealthBench Hard: 42.8%, ahead of GPT-5.4's 40.1%). Notable weakness: coding. Strong token efficiency relative to peers.

On Humanity's Last Exam: 39.9% (thinking mode) vs. 58% (contemplating mode) — a large gap showing the value of its multi-agent inference approach.

Inference-Time Multi-Agent Scaling

An emerging pattern across labs: scaling performance at inference time by orchestrating multiple agents rather than training ever-larger single models.

• Muse Spark's contemplating mode — parallel agents propose and refine solutions, aggregate results, achieve better performance at comparable latency
• Kimi K2.5 (Moonshot AI) Agent Swarm — similar multi-agent inference pattern

The implication: inference-time compute is becoming a tunable lever alongside model scale and training compute. This changes the economics of frontier AI — labs can sell higher-capability tiers (more inference-time agents) as premium offerings.

Small Models Closing the Gap

By early 2026, open-weight models running on consumer hardware began outperforming expectations relative to frontier offerings. Key milestones:

• Gemma 4 (Google, April 2026) — the most capable open-weight models from a US company at time of release
• GLM-5.1 (GLM, April 2026) — a 754B-parameter, 1.5TB open-weight model; highly effective but requires substantial hardware
• Qwen3.6-35B-A3B (Alibaba, April 2026) — a 20.9GB model that runs on a laptop and matched or exceeded some frontier models on specific generation tasks source(https://simonwillison.net/2026/May/19/5-minute-llms/)

The gap between frontier API models and locally-runnable open-weight models is narrowing faster than most practitioners expected. See Model Quantization for the compression techniques enabling local deployment.

Open vs. Closed Weights Tension

Meta's pivot away from open weights is described as "a significant loss for the developer community" by observers who have built on Llama. The tradeoffs:

• Open weights: Enables local deployment, fine-tuning, privacy-preserving use cases, and community innovation. US developer ecosystem built significant infrastructure on Llama.
• Closed API: Enables monetization, quality control, safety oversight, and competitive positioning for enterprise customers.

Meta's stated goal is to compete for business customers alongside OpenAI, Google, and Anthropic. Health specialization and multimodal capabilities align with Meta's product interests (billions of camera-equipped users, health as a top AI query category).

Claude Fable 5 and the Long-Task Frontier (June 2026)

Anthropic released Claude Fable 5 in June 2026 — the same underlying model as Mythos but with added safety safeguards. Fable 5 claimed state-of-the-art on nearly all tested benchmarks, with particular strength in software engineering, knowledge work, scientific research, and vision. Critically, the model's lead over competitors grows with task length and complexity — the longer and more difficult the problem, the wider the margin.

Practitioners report that Fable 5 represents a qualitative step change on par with the November 2025 Claude 4.5 release, especially for extended problem-solving sessions. The model handles significantly more ambitious tasks with less supervision, reaching a point where it becomes "tempting to stop looking at the code at all" — though premature trust remains inadvisable in production source(https://x.com/karpathy/status/2064433620832493812).

Practitioner workflow shift: The most distinctive Fable 5 usage pattern is full task delegation — leaving the model looping for hours or overnight and returning to a finished product, rather than iterating back-and-forth. Mike Krieger (head of Anthropic Labs, who had early access to Mythos-class models) describes ending his workday by briefing the model, then waking up to completed work. When a remote service went down mid-task, Fable 5 wrote a workaround, documented it, and continued autonomously. This delegation-first workflow depends on Fable 5's ability to spawn dozens of subagents reliably without losing context of the main task; Claude Code now supports nested subagents up to five layers deep, where each subagent can spawn further subagents.

The tradeoff practitioners feel most acutely is speed vs. intelligence vs. cost. Fable 5 is noticeably slower than alternatives like Cursor's Composer 2.5 Fast or GPT 5.5, which matters for developers who need rapid iteration cycles. Choosing lower reasoning levels (below High or XHigh) feels counterproductive even when the task doesn't warrant full reasoning budget. Fable medium, however, outperforms Opus xhigh on benchmarks while costing less — suggesting that the right strategy is often to use Fable at moderate reasoning rather than Opus at maximum.

Effort-level efficiency in practice: Early practitioner testing suggests the bias toward high effort is even more misguided than benchmark tables imply. On agentic coding tasks, Fable 5 at low effort outperforms both Opus 4.8 and GPT 5.5 at high effort; at medium effort it beats Opus 4.8 on xhigh and max. Across hours of routine coding — small updates in medium codebases, larger changes in big ones — low effort produced identical solutions to high effort, just faster and cheaper. One useful heuristic: asking the model itself to recommend an effort level per task phase. In one practitioner's project plan, only one of four phases warranted high effort; the rest were fine at medium or low. The implication is that most developer tokens spent on elevated reasoning levels are wasted — the model arrives at the same solution, just slower and at higher cost.

Verification as the bottleneck: As delegation scales, the constraint shifts from writing code to verifying agent output. Krieger gives Fable video captures of its own work so it can catch animation glitches that screenshots would miss — an early example of using the model's own multimodal capabilities to close the verification loop. See Agent Proficiency for more on how human skill is evolving toward verification and orchestration.

Pricing and availability: Fable 5 is approximately 2x the cost of Opus via API (vs. Mythos at 5x). At launch, Fable is included in Claude subscription plans through June 22, 2026, after which Anthropic plans to move it to paid credits until serving capacity scales. SemiAnalysis found that $200/month subscription plans actually deliver well beyond the assumed ~$2,000/month equivalent in API tokens when used for long-horizon coding tasks.

The safety layer is notable: Fable 5 ships with safeguards that are deliberately configured conservatively for launch, erring on the side of over-triggering rather than under-triggering. Anthropic has indicated these will be tuned over time based on real-world feedback. Separately, Anthropic introduced a policy where Fable would covertly degrade output quality for ML/AI-related work — a move that drew sharp backlash and was partially walked back to at least make the restriction transparent rather than hidden. See AI Safety & Interpretability for the broader trend of safety-capability bundling in frontier releases.

An adjacent observation: as frontier model capability improves, demand for software does not plateau — it accelerates. Karpathy describes a Jevon's paradox dynamic where working software "on a tap" expands what practitioners ask for: bespoke single-use apps, project-specific dashboards, massively expanded test suites, automated optimization, and large-scale research projects with custom interfaces. The implication is that frontier models don't reduce overall engineering effort — they shift it toward higher-leverage, more ambitious tasks.

Scaling Philosophy: Data Quality Over Compute

Raw compute capacity is growing rapidly — global AI chip production is on a steep exponential, and Anthropic's ARR alone grew from $9B (end of 2025) to $47B (May 2026), with much of it flowing to compute capex. But the consensus among frontier lab insiders is shifting: compute is no longer the binding constraint on progress. The true bottleneck is the data pipeline — specifically, the ability to generate high-quality training signal for increasingly complex tasks.

The RLVR revolution and its limits. Before OpenAI's o1 model, training data meant text: scraped from the internet, written by paid humans, or synthetically augmented. This was expensive and finite, prompting predictions the industry would run out of data. RLVR changed the economics: for tasks where answers can be verified programmatically (give the model a coding problem, run the output through tests, train on what passes), a single well-designed task yields far more supervision than a single written solution. An ecosystem of companies — Mercor, Datacurve, Mechanize — now builds these verification environments at scale, running LLMs in virtual computers and scoring their outputs deterministically. This pipeline has made frontier models exceptionally strong at verifiable tasks like software engineering.

But RLVR has exposed two structural limitations. First, the marginal cost of creating verification environments scales with task complexity — as tasks surpass human capabilities, building fair and deterministic scoring requires teams of specialists, driving costs up sharply. Second, domains that resist deterministic scoring are correspondingly hard to improve. Writing is the canonical example: reward models and LLM judges tend to optimize for coherence and predictability rather than genuine quality, producing output that is competent but recognizably generic. Fable 5 demolishes coding benchmarks but performs relatively poorly on Vending-Bench, which evaluates open-ended business decision-making — precisely the kind of task where deterministic scoring breaks down source(https://x.com/golokhvastov/status/2066269917386027293).

The real-world feedback hypothesis. One speculative but economically plausible path forward: instead of humans manually designing test environments, train models on direct feedback from reality. Rather than simulating business decisions, have AI agents actually run companies and score on real-world metrics — revenue, growth, churn. A marketing subagent gets evaluated on ad campaign performance; an engineering subagent gets scored on technical issues real users encounter. The company-level reward signal would be sparse (perhaps 1% of AI-run companies crossing a meaningful success threshold), requiring large batch sizes — on the order of 5,000–10,000 microcompanies per quarter. At $25K–$100K per company-run, annual costs would range from hundreds of millions to a few billion dollars, feasible given Anthropic's projected budget source(https://x.com/golokhvastov/status/2066269917386027293).

To bootstrap this, frontier labs could employ small human teams to manage groups of AI agents in real business contexts, documenting decisions and information flows. These records would seed synthetic training pipelines, gradually bringing open-ended executive judgment into the training distribution. The thesis: management is simply another domain to be trained on, following the same trajectory that brought software engineering from "often-work" to "mostly-work" in late 2025.

The convergence argument. Whether LLMs have fundamental out-of-distribution creativity limitations matters less than it might seem. If sufficient compute is available, context limits can substitute for sample efficiency — creativity might emerge when models can simulate a lifetime's worth of experiences in context. If compute is insufficient, the result is a transient period of very capable but not omnipotent AIs: enough to automate most computer work and drive rapid economic growth, but potentially bottlenecked on novel innovation. Either way, the economic flywheel — profitable AI agents funding development of smarter AI — points toward continued rapid capability growth. See AGI Definitions for how labs frame this trajectory.

See Also

• AGI Definitions — how labs define progress and milestones
• AI Safety & Interpretability — safety considerations around frontier model deployment
• Agentic Engineering — multi-agent orchestration patterns
• Agent Proficiency — the human skill of directing AI agents effectively
• Model Quantization — how post-training compression enables frontier model deployment on consumer hardware

Sources

• "Big Pharma Bets Big on AI" — Andrew Ng / deeplearning.ai (newsletter, Apr 2026) (link) — Muse Spark release, Meta's closed-weights pivot, benchmark data, contemplating mode, multi-agent inference trend
• The last six months in LLMs in five minutes — Simon Willison (May 2026) — November 2025 inflection point, model rotation, RLVR-driven coding agent quality, small/open-weight models closing the frontier gap
• "This is a super exciting release - Claude Fable 5…" — Andrej Karpathy (tweet, Jun 2026) — Claude Fable 5 qualitative assessment, SOTA benchmark claims, long-task capability lead, safety safeguard tuning, Jevon's paradox of software demand
• New Claude model - Fable — Ben Tossell / Ben's Bites (newsletter, Jun 2026) — Practitioner Fable 5 experience, speed-intelligence-vibe tradeoff, overnight delegation workflow (via Dan Shipper / Mike Krieger interview), nested subagent spawning, Fable medium vs Opus xhigh cost-performance, ML/AI sabotage policy controversy, subscription pricing analysis
• "If you aren't using models with different effort levels, you're probably wasting tokens, and time" — Morgan (tweet thread, Jun 2026) — Fable 5 low-effort outperforming Opus 4.8/GPT 5.5 at high effort on agentic coding, effort-level selection heuristics, practitioner evidence that routine tasks need less reasoning budget than defaults suggest
• The path from Fable to superintelligence — Goliath (tweet thread, Jun 2026) — Data pipeline as true bottleneck over compute, RLVR ecosystem (Mercor/Datacurve/Mechanize), deterministic scoring limitations, real-world feedback hypothesis for training on business outcomes, economic flywheel to superintelligence, Anthropic ARR growth ($9B→$47B)