Your April invoice came in 40% higher than March, and nobody changed a line of code. That is what happened to teams running production workflows on GPT-5.5 the week OpenAI's May 2026 price adjustment landed - a model that was explicitly sold on token efficiency.
The two-year trend of falling inference costs hit a wall in 2026, and then partially reversed again just this week. Understanding the shape of that reversal tells you something important about how to architect AI features at work right now.
How frontier inference prices actually moved in 2026
The story everyone told through 2025 was simple: costs fall 10x every 18 months, wait and everything gets cheap. Inference cost had declined dramatically - 10 to 100x - from 2023 to 2026, driven by model efficiency improvements, hardware advances, and competitive pricing pressure. That story was true, and then it stopped being true.
Closed-frontier providers - Claude, ChatGPT, Gemini - all started November 2023 between $7 and $30 per million tokens, fell over two years, and reversed in 2026. The reversal had a specific shape. GPT-5.5 is priced at $5 per million input tokens and $30 per million output tokens. Its predecessor GPT-5.4 charged $2.50 input and $15 output.
By raising the per-token cost on both, OpenAI effectively neutralized the efficiency advantage that GPT-5.5 delivered. A user who might have expected a 15 to 20 percent cost reduction from better token efficiency instead faced a 40 percent increase in total spending.
Anthropic played it differently. OpenAI raised both sides aggressively, landing GPT-5.5 at the highest price. Anthropic did not move the sticker at all. Opus 4.7 launched at the same price as Opus 4.6, with the roughly 25% effective increase coming from a new tokenizer that produces more tokens per prompt.
Then, two days ago, OpenAI reversed course on the lower tiers. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens, down from $1.00 and $6.00. Terra drops to $2.00 input and $12.00 output from $2.50 and $15.00. Sol is unchanged at $5.00 input and $30.00 output. The reason, according to OpenAI, is that efficiency gains came from improving the models, the inference systems that run them, and the agentic harness connecting them to tools and context. GPT-5.6 models take a more direct path through work. Better routing keeps hardware productive, optimized production software generates tokens more efficiently, and smarter context management helps agents avoid repeating completed work.
The strategic logic behind Sol staying at $5/$30 while Luna drops 80% is legible. OpenAI's response is to defend the high-volume, cost-sensitive workloads at the Luna and Terra tiers - classification tasks, document extraction, request routing, and long agent loops that represent the bulk of production traffic - while preserving Sol's positioning as the premium flagship.
What open-weight inference actually costs by comparison
The closed-frontier price swings are noisy, but the open-weight baseline has been quietly stable. Llama 4 Scout on Together AI costs $0.18/$0.59 per million tokens against Claude Sonnet 4.6 at $3/$15, which is 16x cheaper on input and 25x cheaper on output for workloads where the quality trade-off holds.
If you go further and self-host: at 1 billion output tokens per month, Claude Sonnet 4.6 costs $15,000. Self-hosted Llama 4 70B on an H200 costs $130. That gap is large enough to restructure a budget. Self-hosting beats managed APIs at roughly 2 to 5 million tokens per day on reserved GPU capacity over a 12-month window.
The capability gap that justified paying frontier prices has also narrowed on specific task types. The coding gap has effectively closed: MiMo V2 Pro, MiniMax M2.7, and DeepSeek V3.2 now sit within striking distance of Opus 4.6 on real-world coding workloads, with MiniMax M2.7 specifically costing roughly 50x less per million output tokens. On reasoning-heavy benchmarks, the story is different. Claude Opus 4.6, GPT-5.4 Pro, and Gemini 3.1 Pro Deep Think retain a meaningful lead on reasoning-heavy benchmarks like GPQA Diamond, Humanity's Last Exam, and frontier math, typically by 3 to 8 percentage points.
The hidden cost: the model you licensed is not fixed
Price is the obvious variable. The less-obvious one is stability. The hardest problem with frontier AI in 2026 is not that prices are going up. They are, and considerably so. The harder problem is that the model you license is not the same week to week. While the model weights are stable, it is the harness around the model - default reasoning effort, system prompts, tool-call routing, cache layers - that is changing.
This harness together with the model is what you are running in production. In other words, the AI system you license is not always the system you think you are running.
Open-weight models work the opposite way. The weights you downloaded in March are the weights you are running in November, but you choose where to run them: someone else's cloud, your own GPUs, or anything in between. That version stability is not free - you own the operational burden - but it is a real and underappreciated property for teams building anything that needs reproducible outputs.
A teammate like Beagle, sitting in Slack and answering questions from structured sources, is less affected by harness drift than a long agentic pipeline is. But for teams building document Q&A, classification queues, or anything with compliance requirements attached, the question of what exactly runs behind the API matters in ways that do not show up in benchmark tables.
Beagle in action#eng-ops, 11:02amHow a team should actually route across the tiers right now
The market is not one thing. It is a stack of tiers with very different economics and very different guarantees. Many teams route the hardest fraction of requests - the complex, agentic, high-stakes 20% - to a closed frontier model, while serving the high-volume, lower-complexity 80% on a cheaper open-weight model. That hybrid captures most of the cost savings without giving up the frontier where it matters, and it keeps you portable if pricing or parity shifts.
A rough decision grid for today's market:
| Workload type | Tier to reach for | Why |
|---|---|---|
| Reasoning, multi-step agents, novel tasks | GPT-5.6 Sol / Claude Opus 5 | 3-8 pt benchmark gap still real |
| Coding tasks, RAG, summarization | Open-weight hosted (Llama 4, DeepSeek V4) | Gap has closed; 16-25x cheaper |
| High-volume classification, extraction | GPT-5.6 Luna or self-hosted open-weight | Luna at $0.20 input now competitive |
| Compliance / air-gapped environments | Self-hosted open-weight only | Physical control, version stability |
| Latency-sensitive, low volume | Closed mid-tier (Claude Sonnet 5) | Reliability, no ops overhead |
Batch APIs - 50% off at Anthropic, OpenAI, and Google - are the easiest structural saving for any agent pipeline that is not latency-sensitive. That discount alone cuts the effective cost of a frontier model to open-weight hosted territory for any work that can tolerate async turnaround.
Frontier model inference cost: common questions
Why did LLM inference prices go up in 2026 after falling for two years?
Token demand appears to be growing much faster than supply. Labs hit compute capacity walls and began charging more for the frontier tier while using efficiency gains to cut lower tiers. GPT-5.5 doubled GPT-5.4's price in April; GPT-5.6 Luna then dropped 80% in July as self-optimized inference kernels reduced serving costs.
When does self-hosting an open-weight model beat paying the API?
Self-hosting beats managed APIs at roughly 2 to 5 million tokens per day on reserved GPU capacity across a 12-month window, assuming steady load. Below that, the GPU reservation cost exceeds API savings. Above it - and for regulated environments that cannot send data off-premises - self-hosting is the correct default.
Are open-weight models good enough for production in 2026?
The catch is not quality. Open-weight models now match or beat closed frontier on most non-reasoning workloads - high-volume chat, RAG, classification, summarization, code completion on well-defined tasks. For complex multi-step reasoning and frontier agentic tasks, closed models retain a measurable but narrowing lead.
What is the real cost difference between Claude and Llama on the same task?
Llama 4 Scout on Together AI costs $0.18/$0.59 per million tokens against Claude Sonnet 4.6 at $3/$15 - 16x cheaper on input, 25x on output
- for workloads where quality trade-offs hold. The break-even depends entirely on whether the quality gap forces more retries or longer prompts to compensate.
Does the model I'm calling via API stay the same over time?
Not reliably. While the model weights are stable, it is the harness around the model - default reasoning effort, system prompts, tool-call routing, cache layers - that is changing. This harness together with the model is what you are running in production. In other words, the AI system you license is not always the system you think you are running. Open-weight self-hosted deployments are the only way to lock weights and harness simultaneously.