The Enterprise LLM Landscape: Benchmarks, Training, Context, Inference Costs, and Model Selection

Updated 1 October 2026. Every model, price and score in this article is as of that date and comes from the sources listed, with vendor-reported results labelled as such. Prices are list prices in US dollars per million tokens. INVISOM has no commercial relationship with any provider named. This article was researched and drafted with AI assistance, including a model from one of the vendors compared, and was reviewed for factual accuracy and balance.

The frontier is a moving tie

September 2026 shows why a static ranking of large language models is of little use to an enterprise. In one month:

• OpenAI released GPT-6 Astra on 3 September, followed by GPT-6 Sol and GPT-6 Luna, then GPT-6.1 Sol at its DevDay on 29 September. SiliconANGLE reported that OpenAI is withholding GPT-6.1 Astra over safety concerns.
• Anthropic released Claude Fable 5.1, then Claude Opus 5.5 on 22 September and Claude Sonnet 5.5. A sibling model, Claude Mythos 5.1, is offered only by invitation through Anthropic's Project Glasswing.
• Google shipped Gemini 3.8 Flash on 2 September, its third Flash release in six weeks, and on 30 September announced Gemini 4 Argon, which is rolling out first to selected cyber defenders.
• Meta released Muse Spark 1.3, a hosted model from its Superintelligence Labs, while DeepSeek and Alibaba's Qwen team published new open-weight models in August and September.

Two patterns stand out. First, all three leading labs now hold their most capable model back from general release. Second, the gaps between the generally available frontier models are small and inconsistent. Each lab leads on some benchmarks and trails on others, and independent leaderboards disagree about who is first.

Our thesis: for most enterprise decisions, "which model is best?" is the wrong question. Differences of a few benchmark points are smaller than the differences created by who runs the test and at what effort setting. Price differences between model tiers, by contrast, run to two orders of magnitude, and the route through which you consume a model decides your data residency, security controls and even which features you get. The durable investments are a deliberate choice of deployment routes, an evaluation pipeline built on your own tasks, and an architecture that lets you swap models at each release.

Three ways to consume a model

Every option sits in one of three deployment models, and the boundaries have blurred.

Proprietary API, direct from the developer. OpenAI, Anthropic, Google, Meta (Muse Spark), Mistral, Cohere, DeepSeek and Qwen all sell API access. This is usually where new models and features appear first.

Managed cloud. The hyperscalers now resell most frontier families under their own contracts, identity and billing. Amazon Bedrock lists GPT-6 and Claude 5.x models alongside Amazon Nova, Mistral, Qwen, DeepSeek, Kimi, Grok and others, but only Google's open Gemma models rather than Gemini. Microsoft Foundry sells GPT-6, DeepSeek V4, Grok, Kimi, Mistral Large 3, Cohere Command A+, Llama and Microsoft's own MAI models, and offers Claude through Anthropic. Claude is available on all three major clouds; Gemini is offered through Google's own platforms. The same model can differ by route: in features, in regional endpoints and in price. Anthropic, for example, documents a 10% premium for regional over global endpoints on Bedrock and Google Cloud. Our comparison of Claude Code and OpenAI Codex shows how much feature parity can vary for the same product by cloud.

Open weights, self-hosted or on a managed endpoint. The open-weight tier is now credible for many enterprise tasks. Stanford's 2025 AI Index found the gap between open and closed models had narrowed from 8% to 1.7% on some benchmarks in a single year. Options include DeepSeek V4 (MIT licence), Qwen3.8-27B (Apache 2.0), OpenAI's gpt-oss (Apache 2.0), Meta's Muse Glimmer 30B (Apache 2.0), Mistral Large 3 and Small 4 (Apache 2.0) and Google's Gemma. Licences differ: Qwen3.8-Flash-Next, for instance, uses a Qwen community licence rather than Apache 2.0. Hardware differs even more. OpenAI states gpt-oss-120b fits on a single 80 GB GPU; DeepSeek V4 Pro is listed at 1.7 trillion parameters, which means a multi-GPU cluster. Running weights on your own infrastructure means prompts never reach the developer, but you take on patching, scaling, safety tuning and a supply-chain review of the weights and serving code, whatever the model's origin.

Context and modality: a million tokens is table stakes

At the frontier, a context window of about one million tokens is now standard: GPT-6 models accept 1.05 million (with a 922,000-token input cap), Claude Fable 5.1, Opus 5.5 and Sonnet 5.5 accept 1 million, as do Gemini 3.8 Flash, Gemini 3.1 Pro, DeepSeek V4 and Muse Spark 1.3. The differences that matter are elsewhere.

• Input modalities. Gemini models accept text, images, video, audio and PDF natively. GPT-6 and Claude models accept text and images. If the workload is video or call audio, this narrows the shortlist before any benchmark.
• Maximum output. It ranges from 64,000 tokens (Gemini, Claude Haiku 4.5) to 128,000 (GPT-6, Claude) and 384,000 (DeepSeek's API), which matters for long reports, code generation and agent transcripts.
• Long-context pricing. OpenAI prices GPT-6.1 Sol prompts above 272,000 input tokens at twice the input rate and 1.5 times the output rate for the whole request. Google charges more for Gemini 3.1 Pro prompts above 200,000 tokens. Anthropic bills Claude 4.6 and later at standard rates across the full window.
• Knowledge cutoff. Disclosed cutoffs are June 2026 for Claude 5.5 and Fable 5.1, 30 April 2026 for GPT-6 Astra and 6.1 Sol, and March 2026 for Gemini 3.8 Flash. Retrieval, not the cutoff, should carry anything that changes.

A large window is a capacity, not a guarantee that the model uses every part of it equally well. Our analysis of RAG versus fine-tuning covers when long context with caching beats retrieval, and when it does not.

What is disclosed about training, and what is not

For the proprietary frontier, the honest answer to "how big is it and what was it trained on?" is not publicly disclosed. None of OpenAI's GPT-6 models, Anthropic's Claude 5.x models, Google's Gemini 3.x models, Meta's Muse Spark or Amazon's Nova 2 models has a published parameter count, training-token count or training-compute figure in the documentation reviewed. Developers do disclose knowledge cutoffs and, in model cards, outline training approaches in general terms.

This is a documented trend. Stanford CRFM's Foundation Model Transparency Index fell to a mean score of 41 in December 2025, a 17-point drop on 2024, with training data still the most opaque area. The 2026 AI Index notes that the most capable models often disclose the least.

Open-weight developers disclose more, though not everything. DeepSeek's V4.1 Flash model card describes its mixture-of-experts design and pre-training on 45 trillion tokens. Qwen discloses 125 billion total and 6 billion active parameters for Qwen3.8-Flash-Next, and OpenAI 117 billion total and 5.1 billion active for gpt-oss-120b. Training-token counts for Qwen3.8 and Muse Glimmer are not disclosed.

Training-compute figures you see for closed models are estimates. Epoch AI, the most widely used source, labels its estimates as accurate within a factor of 3, 10 or 30, depending on confidence. Never put an estimated parameter count or training cost into a business case or risk assessment as if it were fact. For enterprises the useful disclosure questions are practical: licence terms, intellectual-property indemnities, data-use terms, and whether the provider will tell you when a model version changes.

Benchmarks: read them like a referee

Benchmarks are a reasonable way to build a shortlist and a poor way to make a final choice. Four problems recur.

The referee changes the result. On Terminal-Bench 4.0, Anthropic reports 66.4% for Claude Opus 5.5 and Vals AI measures 65.15%. OpenAI's GPT-6 Astra is reported at 57.9% in Anthropic's launch table and measured at 59.6% by Vals AI. Claude Fable 5.1 scores 55.8% in Anthropic's table and 58.08% in Vals AI's run. Vendor tables can also show rivals ahead: Anthropic's own Opus 5.5 table reports GPT-6 Astra leading on Terminal-Bench-Science (64.6% versus 58.7%) and AutomationBench (41.4% versus 40.0%). Google reports 51.3% on AutomationBench for Gemini 4 Argon, measured in Google's own run. These are different runs and should not be merged into one ranking.

Effort settings move scores more than vendors do. Artificial Analysis scores Claude Opus 5.5 at 58 on its Intelligence Index at maximum effort and 51 at medium; GPT-6 Astra scores 53 at maximum and 50 at medium. Any comparison that does not state the effort level, tools and harness is incomplete.

Benchmarks saturate and leak. The 2026 AI Index reports that performance on SWE-bench Verified rose from 60% to near 100% in a year. In February 2026 OpenAI stopped treating it as a frontier measure: its audit of 138 hard problems found 59.4% had flawed tests or task descriptions, and models could reproduce solutions from a task identifier alone, a sign of training contamination. OpenAI recommended SWE-bench Pro instead.

Differences are often within noise. Anthropic reports a standard error of about 2.6 points on Terminal-Bench 4.0 for Opus 5.5. On LMArena's human-preference leaderboard, the top ten models on 30 September sat within 31 points, with Gemini 4 Argon's lead marked preliminary.

Independent aggregators reduce these problems but do not remove them. As of 30 September 2026, Artificial Analysis ranks Claude Opus 5.5 first, while the Vals Index and LMArena both rank Gemini 4 Argon first, a model most enterprises cannot yet buy.

Inference cost: price per task, not per token

List prices span more than two orders of magnitude.

US dollars per million tokens, standard tier, as of 1 October 2026. Sources: OpenAI, Anthropic, Google, Meta and DeepSeek pricing pages. Cloud-route, regional and negotiated prices differ.

To show what this means in practice, we priced three hypothetical monthly workloads at list prices.

The spread is wide: the same long-running agent workload costs about US$78 a month on GPT-6 Luna and about US$7,800 on GPT-6 Astra. But per-token prices understate how costs actually behave, for three reasons.

1. Tokens are not equal. Each vendor tokenizes text differently; Anthropic notes that its tokenizer for Claude 4.7 and later produces about 30% more tokens for the same text than its predecessor. A cheaper per-token price can still produce a larger bill.
2. Reasoning tokens are billed as output, and their volume depends on the effort setting and the model.
3. Agents differ in how many steps they take. On Vals AI's Terminal-Bench 4.0 run, average cost per task was US$9.58 for GPT-6 Astra, US$13.20 for Claude Opus 5.5 and US$16.51 for Claude Sonnet 5.5. Sonnet 5.5 has half Opus 5.5's per-token price but cost more per task in that run.

Measure cost per completed task on your own workload. Also budget for change: Google has announced that Gemini 3.8 Flash prices double in January 2027, Anthropic dropped a planned price rise for Claude Sonnet 5, and OpenAI renamed its priority tier in July 2026. Lifecycles are short too: Anthropic's retirement commitment for Claude Haiku 4.5 runs only to 15 October 2026.

Self-hosting changes the cost model from tokens to capacity: GPUs (owned or reserved), serving software, engineers, monitoring and evaluation. It tends to pay off only with high, steady utilization or when residency leaves no alternative. Model the break-even with your own utilization figures; published hardware or token prices will not do it for you.

A model-selection scorecard

Score each shortlisted model on your own evidence. Hard requirements act as gates: a model that fails residency, licensing or security requirements is excluded whatever its score.

INVISOM framework. Adjust weights by use case: a customer-facing assistant weights latency higher; a regulated workflow weights residency and auditability higher.

The scorecard only works if the evidence behind it is produced the same way for every candidate. That is the job of the evaluation pipeline.

Two design choices make the pipeline durable. Keep the evaluation set and harness independent of any single vendor's tooling, so the comparison is fair. And put a thin routing layer between applications and models, so a better or cheaper model can be swapped in after it passes the suite. The use-case selection discipline that decides what to build should also decide which tasks go into the set.

The Gulf view: residency first, then language

For organizations in Saudi Arabia and the UAE, two questions often narrow the field before capability does.

Where does inference run? Options vary by provider and route. OpenAI offers data residency and inference residency in the UAE for eligible ChatGPT business customers. Claude models can be called on Amazon Bedrock from the UAE and Bahrain regions through global cross-Region inference, which keeps data at rest in the source region while inference may run elsewhere. Anthropic's own API offers US or global inference. Where personal or classified data must be processed in-country, an open-weight model on in-country infrastructure is often the only option that satisfies the requirement. Our NDMO implementation guide covers the classification rules that drive this.

How well does it handle Arabic? The region has its own models: Jais 2, a 70-billion-parameter open-weight model trained from scratch, and ALLaM and Falcon Arabic, built by continued pretraining of existing models. Global frontier models also handle Modern Standard Arabic well in many tasks. Dialects, tone and domain terminology are where differences appear, so Arabic test cases belong in the evaluation set from the start. A separate INVISOM analysis covers Arabic-first models in depth.

What to do in the next 90 days

1. Pick two or three deployment routes, not one model. Typically a managed-cloud default, a direct API for frontier features, and an open-weight option for residency-bound or high-volume work.
2. Build an evaluation set from real work: a few hundred tasks with known good outcomes per priority use case, including Arabic where relevant.
3. Run a structured bake-off across at least two proprietary vendors and one open-weight model, recording model version, date, effort setting and tools for every run.
4. Report cost per completed task and 95th-percentile latency, not price per token, to whoever approves the budget.
5. Put a routing layer in place so applications call a capability, not a model name, and pin model versions in production.
6. Get residency, retention and licence terms in writing for each route, model by model.
7. Set a re-evaluation trigger: every major model release, price change or retirement notice, and at least quarterly.

Frequently asked questions

Which LLM is best for enterprises in 2026?

No single model is best. As of 1 October 2026, Artificial Analysis ranks Claude Opus 5.5 first, while the Vals Index and LMArena rank Google's Gemini 4 Argon first, and GPT-6 Astra leads some vendor-reported benchmarks. Choose by testing on your own tasks, cost per task and deployment route.

How much do enterprise LLMs cost per million tokens?

As of 1 October 2026, list prices range from US$0.10 input and US$0.50 output (GPT-6 Luna) to US$10 input and US$50 output (GPT-6 Astra, Claude Fable 5.1). Cached input, batch discounts, long-context surcharges and reasoning tokens change real costs substantially.

Are open-weight models good enough for enterprise use?

For many tasks, yes. Stanford's AI Index found the open-closed gap narrowed to 1.7% on some benchmarks. DeepSeek, Qwen, OpenAI's gpt-oss, Meta's Muse Glimmer and Mistral all publish open-weight models. Check licences, hardware needs and your own evaluation results.

Why can't we rely on published LLM benchmarks?

Scores depend on who runs the test, the effort setting and the harness, and popular benchmarks saturate or leak into training data. OpenAI stopped using SWE-bench Verified in 2026 after finding flawed tests and contamination. Use public benchmarks to shortlist, then test on your own tasks.

Sources and further reading

• OpenAI: API pricing, GPT-6.1 Sol model page, Why SWE-bench Verified no longer measures frontier coding capabilities (February 2026), gpt-oss-120b model card. Accessed 1 October 2026.
• OpenAI's GPT-6.1 Sol delivers Astra-like performance at a dramatically lower price, SiliconANGLE, 29 September 2026.
• Anthropic: Claude models overview, pricing, data residency, Introducing Claude Opus 5.5 (22 September 2026). Accessed 1 October 2026.
• Google: Gemini API pricing, Introducing Gemini 3.8 Flash (2 September 2026), Gemini 3.8 Flash model card, Gemini 4 Argon (30 September 2026).
• Meta: Muse Spark 1.3, Introducing Muse Spark (8 April 2026); Muse Glimmer on Hugging Face (10 August 2026).
• Other providers: Mistral models, Cohere models, Amazon Nova, DeepSeek pricing, DeepSeek-V4.1-Flash, Qwen3.8-Flash-Next.
• Cloud catalogues: Amazon Bedrock models at a glance; Foundry Models sold by Azure, Microsoft.
• Independent evaluation: Artificial Analysis leaderboard; Vals AI Terminal-Bench 4.0 (29 September 2026); LMArena text leaderboard (30 September 2026).
• Research: Stanford HAI, 2026 AI Index takeaways (13 April 2026); 2025 AI Index Report; Foundation Model Transparency Index (December 2025); Epoch AI, Data on AI models.
• Gulf: OpenAI data and inference residency; AWS Bedrock Middle East cross-Region inference (24 February 2026); Jais 2 release (9 December 2025).