Which LLM Should You Use?

A practical comparison of frontier language models for business applications. Cut through the marketing and find the right model for your specific needs.

Last updated: October 2026

Model Comparison

Model	Context	API Pricing (in/out)	Consumer Access
Claude Fable 5.1 Anthropic	1M tokens	$10 / $50 per 1M tokens	Claude Pro $20/mo (Pro/Max/Team/Enterprise)
Claude Opus 5.5 Anthropic	1M tokens	$4 / $20 per 1M tokens ($0.20/M cache reads)	Claude Pro $20/mo (Pro/Max/Team/Enterprise)
Claude Haiku 5.5 Anthropic	1M tokens (128K max output)	$0.10 / $0.50 per 1M tokens up to 100K-token prompts; $0.50 / $2.50 above ($0.01/M cache reads)	API-first (Claude API, Bedrock, Google Cloud, Microsoft Foundry)
GPT-6 Astra OpenAI	1.05M tokens	$10 / $50 per 1M tokens (2x input/cache, 1.5x output above 272K input tokens)	ChatGPT Plus $20/mo (staged rollout to all paid tiers after Sept 3 launch)
GPT-6.1 Sol OpenAI	1.05M tokens (128K max output)	$2 / $10 per 1M tokens (higher rate above 272K input tokens)	ChatGPT Plus $20/mo
GPT-6 Luna OpenAI	1.05M tokens (128K max output)	$0.10 / $0.50 per 1M tokens ($0.01/M cached input) - permanent pricing	Free tier (unlimited text chats) and ChatGPT Go $8/mo default; also ChatGPT Plus $20/mo
Gemini 3.1 Pro Google	1M tokens	$2 / $12 per 1M tokens (up to 200K; $4/$18 above)	Google AI Pro $19.99/mo (Deep Think needs AI Ultra $99.99/mo)
Gemini 3.8 Flash Google	1M tokens (64K max output)	$0.75 / $3.75 per 1M tokens (intro rate through Dec 31, 2026; $1.50/$7.50 from Jan 1, 2027)	Google AI Pro $19.99/mo; runs on the free Gemini app tier
GLM-5.3 Z.ai	1M tokens	$1.40 / $4.40 per 1M tokens (official GLM-5.3 rate; $0.26/M cached input)	GLM Coding Plan from $12.60/mo (annual) or $18/mo; 753B-param open weights released Aug 28, 2026 under a bespoke license, not MIT
Grok 4.7 SpaceXAI (formerly xAI)	500K tokens	$2 / $6 per 1M tokens (<200K prompt); $4 / $12 above 200K ($0.50/M cached input below 200K, $1/M above)	SuperGrok $30/mo (Heavy $300/mo for top rate limits)
Muse Glimmer Meta	131K tokens	Self-hosted / ~$0.30–$1.20 per 1M via third-party hosts (no official Meta API price)	Open weights (Apache 2.0)
DeepSeek V4 Pro DeepSeek	1M tokens	$1.32 / $3.96 per 1M peak, $0.66 / $1.98 off-peak (up from $0.435/$0.87 flat)	API only

Strengths & Limitations

Claude Fable 5.1

Anthropic

Strengths

Released Sept 1, 2026 as the successor to Fable 5, at the same input/output price with cache reads cut to a quarter of the cost ($0.25 vs $1 per 1M)
Capability gains concentrate in six areas: long-running agentic coding, document/spreadsheet/slide creation, multistep research, dense-document vision, full-context reasoning, and computer use
Full 1M context at standard pricing, with adaptive thinking always on and a five-level effort toggle, including mid-conversation effort changes
Output carries a statistical text watermark and generated files carry signed C2PA content credentials

Limitations

Priciest Anthropic model - 2.5 times Opus 5.5's per-token cost, and Anthropic itself recommends starting with Opus 5.5 for most workloads
Three breaking API changes versus Fable 5: forced tool use is unsupported, thinking blocks are tied to this model, and editing earlier turns in a conversation can invalidate them
Batches fewer tool calls per turn than Fable 5 in agent loops, which can add round-trips and latency for the same task
Claude Mythos 5.1 (same capabilities) remains invitation-only through Project Glasswing - not broadly available

Claude Opus 5.5

Anthropic

Strengths

Released Sept 22, 2026, replacing Opus 5 at $4/$20 per 1M tokens - a 20% price cut - while matching Claude Fable 5.1 on most work
89.9% on SWE-bench Pro and 66.4% on Terminal-Bench 4.0 at xhigh effort, both well above Opus 5's prior scores
Costs roughly 40% less to run than Opus 5 at default effort, and about 60% less than Fable 5.1, for comparable agentic performance
First Opus model to ship with Fable 5.1-level safeguards for cybersecurity, biological, and model-distillation risk

Limitations

Still trails Fable 5.1 on the hardest reasoning and longest-horizon agentic work - Anthropic's guidance is to move up a tier when Opus 5.5 falls short at higher effort
Fast Mode (research preview) costs $8/$40 per 1M tokens - a real premium over standard pricing, though cheaper than Opus 5's old $10/$50 Fast Mode rate
Anthropic's own evaluations still put Mythos-class models ahead on cybersecurity tasks
Launched the same week as GPT-6 Sol/Luna and Grok 4.7, so independent head-to-head benchmarking is still thin

Claude Haiku 5.5

Anthropic

Strengths

Released Oct 7, 2026 as Anthropic's fastest and lowest-priced current model, aimed at classification, routing, extraction and subagent work
Context window grows to 1M tokens and output to 128K, up from 200K and 64K on Haiku 4.5
Adaptive thinking with an effort control (default medium), and text and image input
Minimum cacheable prompt drops to 512 tokens from 4,096 on Haiku 4.5, so short repeated prompts can use the $0.01/M cache-read rate

Limitations

Prompts over 100K tokens are billed at five times the base rate on every token in the request, so the 1M window is not cheap at full length
Uses the newer tokenizer: the same text counts as roughly 30% more tokens than on Haiku 4.5, which narrows the real saving against the headline price
Breaking API changes from Haiku 4.5: sampling parameters, assistant prefill and manual thinking budgets now return errors, and safety classifiers can return a refusal with no server-side fallback
No independently verified benchmark scores were available at refresh time; Anthropic's pitch is speed and cost, not frontier reasoning

GPT-6 Astra

OpenAI

Strengths

Launched Sept 3, 2026 as OpenAI's new flagship, replacing GPT-5.6 Sol across ChatGPT and the API
Saturates several OpenAI benchmarks: ~97.6% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench (versus Sol's 78.5%)
Computer use: 72.6% on OSWorld 2.0 in roughly 40 minutes per task, versus GPT-5.6 Sol's 65.7% in roughly 75 minutes
Native text-and-image multimodal input across the full 1.05M token context window

Limitations

First OpenAI model to cross the "critical" cybersecurity threshold under its Preparedness Framework - offensive-capability access is gated behind the vetted "Daybreak" program, and the general public version refuses those tasks
Rollout was staged and rocky: launched to a limited group Sept 3, with Sam Altman apologizing for a "messy rollout"; users also reported it burning through ChatGPT usage quotas roughly 3-6x faster than GPT-5.6 Sol because of its cost structure and Fast Mode multiplier
Trails Claude Fable 5.1 on Humanity's Last Exam with tools (57.2% vs 65.0%), so it isn't a clean sweep
A GPT-6.1 Astra was reportedly planned for DevDay on Sept 29 but shelved over safety concerns (Gizmodo reporting; not confirmed on an OpenAI page we could retrieve), so GPT-6 Astra remains the top GPT-6 model
OpenAI's own system card notes reduced chain-of-thought monitorability alongside claimed gains in alignment and jailbreak robustness
OpenAI's Sept 16, 2026 Misalignment Reporting Framework disclosed that an unreleased Astra-family training checkpoint had inserted fabricated jailbreak-style instructions into its own training-run summaries to get successor models to disregard developer messages; OpenAI says this happened in an internal run separate from the model that shipped

GPT-6.1 Sol

OpenAI

Strengths

Launched Sept 29, 2026 at OpenAI DevDay as an upgrade to GPT-6 Sol (API ID gpt-6.1-sol), at the same $2/$10 price as the model it replaces
Priced at a fifth of GPT-6 Astra's per-token rate; OpenAI's pitch, as reported, is near-Astra intelligence at that price
Full 1.05M token context window and 128K max output, the same as Astra
Arrives after the roughly 50% GPT-6 API price cut that launched with GPT-6 Sol and Luna on Sept 22

Limitations

Third-party benchmark data is thin and comes mostly from aggregators; we have not verified any score against an OpenAI page, so none is quoted here
ChatGPT rollout was staged: one report had it in ChatGPT Work and Codex but not yet in Chat at launch, and sources conflict on which tiers have it now
OpenAI's documentation still lists GPT-6 Sol alongside 6.1, and sources disagree on which is the default mid-tier ID - confirm gpt-6.1-sol against the live models page before pinning
Below flagship GPT-6 Astra on the hardest multi-step reasoning and agentic work

GPT-6 Luna

OpenAI

Strengths

Launched Sept 22, 2026, cutting GPT-5.6 Luna's already-low price a further 50% on input and 58% on output
Reported as the default model for ChatGPT Free and Go users from Oct 8, 2026, though one report says the help centre still lists GPT-5.6 Luna for those tiers - treat as contested
66.6% on DeepSWE (~$0.22 per task) and 50.9% on Agents' Last Exam (~$0.15 per task) - strong cost-efficiency for an entry-level model
Full 1.05M token context window, same as Sol and Astra

Limitations

Gap to GPT-6.1 Sol and flagship GPT-6 Astra widens significantly on the hardest reasoning and long-context tasks
Not suited to nuanced creative or deep analytical work

Gemini 3.1 Pro

Google

Strengths

Adjustable reasoning depth (Low/Medium/High) — High mode runs as a "Deep Think Mini"
77.1% on ARC-AGI-2, leading abstract reasoning benchmarks
Natively multimodal: text, images, video, and audio
Strong agentic tool-use with fewer wasted tool calls

Limitations

Pricing doubles above 200K tokens ($4/$18 per 1M)
Still labelled preview as of August 2026, with some output inconsistency pre-GA
Flatter creative/emotional tone versus earlier Gemini releases

Gemini 3.8 Flash

Google

Strengths

Replaced Gemini 3.7 Flash on Sept 2, 2026 - Google's third Flash-tier release in six weeks (3.6, 3.7, 3.8)
Terminal-Bench 2.1 jumped to 90.8% from 3.7 Flash's 81.6%; Artificial Analysis Intelligence Index up 3 points to 59
Gains concentrate in agentic work - tool use and terminal coding, with the Agentic Index up roughly 11% relative to 3.7 Flash - while raw coding scores are close to flat
Carries over 3.7 Flash's introductory pricing rather than resetting the discount clock with each release

Limitations

Introductory pricing still expires Jan 1, 2027, when both input and output rates double to $1.50/$7.50
Max output still capped at 64K tokens
A separate '3.8 Flash Cyber' variant with stronger security capabilities is gated to trusted government and enterprise customers, not generally available
Not ideal for long-form nuanced writing or deep analytical work

GLM-5.3

Z.ai

Strengths

Released Aug 14, 2026 on the same GLM-5.2 base model, with every capability gain from a 10x expansion of long-horizon post-training
Largest gains on long-horizon coding: Terminal-Bench 3.0 jumped from 4.6 to 28.3 and DeepSWE v1.1 from 46.2 to 66.9
Rolled out immediately to all existing GLM Coding Plan subscribers, with API access live from day one
Open weights (753B-param MoE) now live on Hugging Face — free to download, fine-tune, redistribute and use commercially for virtually all users

Limitations

Weights ship under a bespoke "GLM-5.3 License," not the MIT license GLM-5.2 used — Model-as-a-Service operators above $10B/12-month revenue need a Z.ai security review before commercial use
Post-training produced an offensive-security capability jump the company says outgrew its own expectations, surfacing 1,097 critical/high-severity vulnerabilities in evaluated software
Trails GPT-5.6 Sol by 6.3 points on Terminal-Bench 3.0 (34.6 vs 28.3) and 5.8 points on DeepSWE v1.1 (72.7 vs 66.9)
Z.ai-hosted API still routes through China-based infrastructure — a concern for regulated or sensitive workloads

Grok 4.7

SpaceXAI (formerly xAI)

Strengths

Released Sept 21, 2026 - xAI's best model yet, "a notable improvement over Grok 4.6 at the same price and speed"
2.1T-parameter base model, up 40% from Grok 4.6's 1.5T, trained with a longer reinforcement-learning run weighted toward hard, hours-long tasks
46.3% on CursorBench 4.0 (up from 4.6's 40.4%) and 71.0% on DeepSWE v1.1 at high effort (up from 65.2%)
AA-Omniscience hallucination rate improved to 29% from 34%, lifting the overall knowledge-reliability index from 30 to 32

Limitations

Still fabricates an answer more than one time in four when it doesn't know one, despite the improved hallucination rate
Context window unchanged at 500K, smaller than the 1M+ windows most rivals now offer
Pricing held flat versus Grok 4.6 despite the capability gain, so its per-token cost advantage over rivals hasn't widened
Musk has sketched Grok 4.8 (2.5T params, still training as of Sept 13), a "4.9" pitched as "Astra/Fable class," and Grok 5 as a possible frontier leader - none has a model ID, price, or benchmarks yet

Muse Glimmer

Strengths

Released Aug 10, 2026 under Apache 2.0 — Meta's most permissive license yet, with no commercial-use or EU restrictions (unlike Llama 4's license)
29.6B-parameter dense model distilled from Meta's Muse Spark, natively multimodal (text and images) and tuned for local agent and coding workflows
Runs on consumer hardware — a 17GB 4-bit quant fits a single consumer GPU or Mac
Replaces Llama 4 Maverick as Meta's current open-weight offering

Limitations

Much smaller context window (131K) than Llama 4 Maverick's 1M
Meta's actual flagship is now the closed, paid Muse Spark 1.3 (released Sept 2, 2026, succeeding 1.2) — same $1.25/$4.25 per 1M price, or $0.10/$0.20 on a data-sharing "Contributor" tier; Glimmer is a distilled, smaller sibling, not directly comparable on capability
Meta has still not open-weighted any Muse Spark generation, despite saying back on Aug 10, 2026 it would do so "soon" — no date or license terms confirmed as of this refresh
No official Meta-hosted API — total cost depends on self-hosting or third-party providers

DeepSeek V4 Pro

DeepSeek

Strengths

Reached general availability on July 20, 2026, after a preview period that began April 24
80.6% on SWE-bench Verified — within 0.2 points of Claude Opus 4.6, at a fraction of the output price even after the August price rise
1.6T-parameter Mixture-of-Experts architecture (~49B active params), 1M token context window, unchanged from preview
Open weights available under MIT for self-hosted deployment; the new DeepSeek-V4.1-Flash sibling (released Sept 10, 2026) undercuts the prior V4-Flash tier at $0.30/$1.20 peak, $0.15/$0.60 off-peak per 1M tokens

Limitations

Data stored on China-based servers — significant privacy risk
New peak/off-peak surge pricing took effect Aug 16, 2026, roughly tripling to quadrupling rates versus the prior flat price, with peak hours running 01:00–04:00 and 06:00–10:00 UTC
Off-peak rates are still higher than the old flat rate in every category — a price increase with a time-of-day discount, not a genuine cut
Legacy deepseek-chat and deepseek-reasoner model aliases were retired on July 24, 2026 — calls using those names now fail
DeepSeek proposed rerouting V4 Pro billing to cheaper V4.1-Flash rates from Sept 14, 2026, then reversed the plan on Sept 11 "in response to user demand" — V4 Pro pricing and billing above are confirmed unchanged

Use Case Recommendations

Different tasks demand different trade-offs. Here are Agathon's recommendations based on common business scenarios.

Use Case	Recommended	Alternatives	Notes
Complex Analysis & Research	Claude Fable 5.1	GPT-6 AstraGemini 3.1 Pro	When accuracy and depth matter more than speed or cost
Production Applications	Claude Opus 5.5	GPT-6.1 SolGLM-5.3	Anthropic's new default for quality-sensitive production workloads
Long Document Processing	Gemini 3.1 Pro	Claude Fable 5.1GPT-6 Astra	Native 1M context; watch the pricing step above 200K tokens
Reasoning, Math & Science	Gemini 3.1 Pro	Claude Fable 5.1GPT-6 Astra	Deep Think mode leads on abstract reasoning (77.1% ARC-AGI-2)
Customer Service & Chatbots	Gemini 3.8 Flash	Claude Haiku 5.5GPT-6 Luna	Half price through end of 2026 on an introductory rate; handles complex queries with tool use
Budget-Conscious Projects	DeepSeek V4 Pro	GLM-5.3GPT-6 Luna	Near-frontier performance well below frontier prices, even after the August price rise
On-Premise / Air-Gapped	Muse Glimmer	GLM-5.3 (self-hosted)DeepSeek V4 Pro (self-hosted)	When data cannot leave your infrastructure; Apache 2.0 licensed with no EU restrictions, and light enough for a single GPU
Creative Writing	Claude Fable 5.1	Claude Opus 5.5GPT-6 Astra	Anthropic's top model for nuanced, long-form writing
Code Generation	Claude Opus 5.5	GLM-5.3Grok 4.7	Strong coding performance at well under half of Fable 5.1's price; GLM-5.3 and Grok 4.7 for cheaper agentic coding
Multimodal (Images/Documents/Video)	GPT-6 Astra	Gemini 3.1 ProGrok 4.7	Broadest native multimodal understanding across formats and file types
Real-Time & Social Context	Grok 4.7	GPT-6 AstraGemini 3.8 Flash	Native access to real-time X data for current-events tasks via SpaceXAI's X integration

Choosing a model is the easy part

This table changes every month, and so does the model under any AI product you rely on. If you're investing in an AI company, or committing serious budget to one, the harder questions are what's actually been built, how much of it rests on someone else's model, and whether it will still hold up after the next update. That's the independent assessment I do, and I take no build work on anything I assess.

Talk to me before the money moves How AI due diligence works Why independence matters

* Pricing reflects October 2026 rates and may change. Check provider websites for current pricing.

* GPT-6 Astra's $10/$50 per 1M token pricing rises to 2x input/cache and 1.5x output above 272K input tokens; ChatGPT and API access were still completing a staged rollout as of this update.

* Model capabilities and context windows are based on publicly available documentation.

* Recommendations reflect Agathon's experience across client engagements. Your specific requirements may differ.

The Agathon Briefing

A regular note with two or three thoughts on seeing the whole AI system clearly, people included, from someone who assesses those systems for a living. Written by hand, read in five minutes.

No more than one briefing a month