LLM 2022–2026: Falling Costs and Corporate Agents

In four years, language models have changed in nature and scale. Token prices have collapsed, architectures like MoE have emerged, and agents are leaving the lab for the field. Adoption has shifted towards enterprises, with Anthropic leading in 2025, while DeepSeek has imposed its rapidly paced innovations.

In 2025, adoption shifts towards enterprises and agents

After a year in 2024 dominated by the general public, adoption moved towards enterprises in 2025. The Claude 4 generation, released in May 2025, brought extended reasoning capabilities associated with tool usage, allowing the model to think and execute actions in a loop. Claude Code, initially offered as a research preview, achieved $1 billion in annualized revenue within six months and, according to SemiAnalysis, this tool accounted for about 4% of public commits on GitHub. By September, Claude Sonnet 4.5 became the first model considered by companies to be reliable enough to operate autonomously for over 30 hours, with the ability to revert to previous checkpoints and manage background sub-agents.

Meanwhile, the Model Context Protocol, presented as an open standard to connect enterprise toolchains to AI, established itself as an interoperable foundation, reaching nearly 97 million installations by March 2026. Adoption figures illustrate this shift: over 300,000 professional clients in October 2025, nearly 80% of revenue generated by enterprises, 470,000 positions deployed at Deloitte, and revenue growth from $1 billion to $9 billion in twelve months.

Reasoning, extended contexts, and falling token prices

This period saw a qualitative leap in reasoning. In September 2024, OpenAI delivered o1, trained to "think before speaking" through reinforcement on thought chains, priced at $15/60 per million tokens. In terms of capabilities, Gemini 1.5 Pro extended the context window to one million tokens in February 2024, while in May 2024, GPT-4o combined reactive voice and vision within a single model. The year 2025 saw the arrival of models such as Gemini 2.5 and descendants of the GPT-4o range, natively trained on video, sound, and interfaces, with live video understanding moving from demonstration phase to product phase.

This power surge was accompanied by a marked decrease in prices. By the end of 2023, GPT-3.5 Turbo was priced at $0.50/1.50 per million tokens and GPT-4 Turbo at $10/30, following an intermediate price of $10/30 at the launch of GPT-4 Turbo with 128K context and knowledge cut off in April 2023. Initially, at the end of 2022, access to the best model via API was around $20 per million tokens, and GPT-4 at launch cost $30/60.

The DeepSeek turning point: massive MoE and optimized training

DeepSeek accelerated the architectural shift towards Mixture-of-Experts. DeepSeekMoE proposed a model with a total of 16 billion parameters, with only 2.7 billion active per token. In February 2024, DeepSeekMath introduced Group Relative Policy Optimization, a more economical reinforcement method suited for multi-step reasoning. DeepSeek-V2, in May 2024, pushed the approach to 236 billion total parameters, 21 billion active, a 128K token window, and Multi-head Latent Attention to compress the KV cache. In December 2024, DeepSeek-V3 expanded the concept to 671 billion total parameters and about 37 billion active per token, with shared experts accessible to all tokens. Other contributions included Multi-Token Prediction, finely granular expert routing, and large-scale training in FP8.

The trajectory began as early as 2023: the lab was founded in July by Liang Wenfeng, DeepSeek Coder was published on November 2, followed by DeepSeek LLM 7B/67B on November 29, competent models of the Llama 2 class but little noticed outside of China. The described strategy combines technical publication, weight release, and aggressive price cuts. In the open-source realm, Mixtral 8x7B had laid the groundwork in December 2023 with 8 experts and about 47 billion total parameters for 13 billion active per token.

From demos to use cases: agents, tools, and protocols

The year 2024 was marked by the first compelling demonstrations of agents: in October, Anthropic unveiled "Computer Use," which allows Claude to manipulate a screen; "Devin" was presented as an autonomous agent for software engineering; OpenAI introduced "Operator" in January 2025. Methodologically, RLHF ceded more ground to RL with verifiable rewards, where objectives consisted of unit tests, mathematical solutions, and formally validated outputs. Feedback processes became faster and more automated, with humans taking on more of a coaching role than that of a corrector.

In enterprises, this agentic approach was accompanied by control and orchestration mechanisms: Claude 4 combined deliberation and tooling, Claude Sonnet 4.5 was used autonomously over long durations, and the Model Context Protocol became established to link toolchains to AI.

Milestones of 2023 and initial context of 2022

The year 2023 laid the technical foundations: the launch of GPT-4 in March (multimodal, text then vision), function calls in June to connect tools via JSON, and Code Interpreter in July for Python execution. In November, GPT-4 Turbo brought a 128K token window, knowledge cut off in April 2023, and a price of $10/30. Google responded in December with Gemini 1.0, available in Ultra, Pro, and Nano versions, and RAG became the standard method for injecting relevant data into context. The open-source landscape was structured around Llama 2 (open weights, 7 to 70B), Mistral 7B which surpassed larger models, and the popularization of MoE. Despite this, open models remained behind the state of the art while offering a good cost-benefit ratio.

The starting point, at the end of 2022, was the launch of ChatGPT on November 30, free to the public and built on GPT-3.5, a descendant of GPT-3 (175B parameters) whose exact architecture has not been published. API access to the best model then cost about $20 per million tokens, and ChatGPT reportedly reached 100 million users within two months. By the end of 2023, prices had fallen to $0.50/1.50 for GPT-3.5 Turbo and $10/30 for GPT-4 Turbo.

These milestones are set against a tumultuous 2022 context: the war in Ukraine triggered in February, sanctions and energy volatility; the Fed's monetary tightening to 4.25–4.50% in December with a 10-year yield around 3.88%; a dollar at 131 yen, U.S. debt at $31 trillion, about $140 billion in cumulative debt for four software giants, and Brent at $93.