AI Infrastructure Needs to Be Model-Agnostic

If you're making decisions about AI for your business, you're probably asking the wrong question.

The question most organisations ask is: "Which AI provider should we use?" They evaluate OpenAI, Anthropic, and Google, pick one, and build everything around it.

That made sense two years ago when one provider was clearly better than the rest. It doesn't make sense anymore. The data from the past two years shows three things that should change how every business thinks about AI infrastructure:

1. There is no consistent "best" provider. The #1 position has changed hands eight times in two years -- roughly every 3 months. OpenAI's GPT-5.5 just retook the lead in April 2026.
2. Open source keeps pace with proprietary. The best open-source models now match paid models in intelligence -- at 3-23x lower cost.
3. Frontier AI saw a 99% cost decline. What cost $45 per million tokens in 2023 costs $0.27 today.

The implication is clear: the winning strategy isn't choosing the best vendor. It's building infrastructure flexible enough to use any of them -- and switch when it makes sense.

Here's the full picture.

1. The "Best" AI Provider Changes Every Few Months

This is the chart that should end the "which vendor should we pick" conversation.

It tracks each of the six frontier AI providers -- OpenAI, Anthropic, Google, xAI, DeepSeek, and Moonshot -- and what their best model scored on the Artificial Analysis Intelligence Index v4.0 at each point over the past two years. The shaded area is the leading score (SOTA) at each moment.

In March 2024, Anthropic was #1 with Claude 3 Opus. By December, OpenAI took the lead with o1. xAI briefly nudged ahead in February 2025 with Grok 3 mini (high). Google overtook the field in March with Gemini 2.5 Pro. OpenAI reclaimed it in April with o3. xAI grabbed SOTA again in July with Grok 4 -- the first time a non-Big-Three provider had clearly led. OpenAI took it back in August with GPT-5. Anthropic retook the top in November with Claude Opus 4.5. OpenAI in December with GPT-5.2. Google in February 2026 with Gemini 3.1 Pro. OpenAI tied through March -- and then in April, OpenAI pulled ahead again with GPT-5.5 at 60.

Ten lead changes across four different providers in two years. No one has held the #1 position for more than a few months.

And the chart's other story is even more important: by April 2026, all six providers are within 8 points of frontier. DeepSeek (52) and Kimi (54) -- both open-weights -- are now closer to GPT-5.5 than Claude 3 Opus was two years ago. Meanwhile, SOTA itself has improved 3.3x in two years. The frontier is advancing fast, and there are now six labs in striking distance of it.

If you signed a multi-year contract with one vendor because they had the best model at the time, there's a good chance they've already been overtaken. And you're locked in regardless.

2. Open Source Keeps Pace with Proprietary

While the proprietary vendors have been trading the lead among themselves, something more significant has been happening beneath the surface: open-source AI has closed most of the gap.

This chart shows how the best open-source model compares to the best proprietary model over time, using independent benchmark scores from Artificial Analysis:

Two years ago, the best open-source model was roughly half as capable as the best proprietary one. Today the gap is about 10% -- GPT-5.5 sits at 60, and Moonshot's Kimi K2.6 (released 20 April 2026) is the new open-source leader at 54. Open-source models improved 4.5x in capability over this period, compared to 3.3x for proprietary.

Where Things Stand Today

Here's the current scoreboard -- the best models from each side, grouped by capability tier:

At the upper tier, the race is essentially over. Models like MiniMax-M2.7 and Qwen 3.6 Plus -- both free to use commercially -- perform within 12% of Gemini 3.1 Pro and GPT-5.4. For the vast majority of business applications, that difference is imperceptible.

The only place proprietary AI leads meaningfully is at the absolute frontier -- where GPT-5.5 just opened up a fresh lead at score 60. Historically, every time a new proprietary frontier model launches, open source closes the gap within 6-9 months. Expect the same pattern here.

3. Intelligence Matched, Wildly Different Price

Intelligence is close across tiers. Cost diverges dramatically -- proprietary is 3x to 23x more expensive for the same intelligence.

This chart shows how much intelligence each model delivers per dollar spent. Think of it like comparing fuel efficiency across cars: which model gives you the most capability for your money?

The most cost-efficient open-source model delivers 94x more intelligence per dollar than the most expensive proprietary option (now GPT-5.5 at $17.50/M blended). Even comparing only the top tier, open-source models offer 10-15x better value than their proprietary equivalents.

To put it plainly: Claude Opus 4.6 and MiniMax-M2.7 are separated by about 6% in capability. They're separated by 20x in price.

4. What Was Frontier Becomes Free -- a 99% Cost Decline

This pattern isn't new. It's been playing out on repeat for three years -- and accelerating.

When GPT-4 launched in March 2023, it cost $45 per million tokens and nothing else came close. Within a year, OpenAI's own GPT-4o cut that to $6.25. Within two years, open-source alternatives matched GPT-4's intelligence for under $1. Today, Gemma 4 delivers twice GPT-4's intelligence for $0.27 -- a 99% cost decline in three years.

Every intelligence tier follows this pattern: a proprietary model launches at a premium, competition drives the price down, and within 12-18 months an open-source model matches it for a fraction of the cost. Whatever you're paying for AI today, you'll likely pay 50-70% less for the same capability within a year.

5. Proprietary Prices Have Stopped Falling

Here's the twist that makes all of this urgent: while AI capability keeps getting cheaper through open-source competition, the prices that proprietary vendors charge have actually been rising.

Through 2023-2024, prices crashed. OpenAI cut by 86%. Google slashed by 55%. It looked like AI was heading towards near-zero cost.

Then the direction changed. Each successive version of GPT-5 has been more expensive than the last: $5.63, then $7.88, then $8.75 -- and now GPT-5.5 launched at $17.50/M blended ($5 input / $30 output), a sharp jump from GPT-5.4's $8.75. OpenAI also introduced a new GPT-5.5-pro tier at $105/M blended ($30 input / $180 output) -- the highest list price for any flagship LLM since Claude 3 Opus in early 2024. Google's Pro line doubled from its floor back to $7. Anthropic held its top price at $45 for over a year before cutting to $15, where it's stayed since.

The open-source lines sit at the bottom of the chart, barely visible -- all below $1, with capabilities that have been rising every generation.

If you've built your AI operations around a single proprietary vendor, you're subject to their pricing decisions. And recent history shows those prices are going up, not down -- even as equivalent capability becomes available for a fraction of the cost elsewhere.

6. Speed Isn't a Differentiator Anymore

One assumption that hasn't held up: that open-source models are slower.

Several open-source models now match or exceed proprietary ones on throughput. Models using Mixture-of-Experts architectures -- which only activate a fraction of their parameters per request -- are especially fast. The speed gap, like the intelligence gap, has largely disappeared.

Where Proprietary AI Still Wins

This isn't a one-sided story. Proprietary models maintain real advantages in specific areas:

The absolute frontier. GPT-5.5 just took the lead at 60 on the intelligence index, with Gemini 3.1 Pro and GPT-5.4 close behind at 57. The best open-source model -- Kimi K2.6, released in the same window as GPT-5.5 -- now sits at 54. For the hardest reasoning, coding, and scientific tasks, that 10% gap still matters, but proprietary's lead was answered almost immediately on the open side rather than going unchallenged.

Adaptive reasoning. Claude and OpenAI's o-series models can dynamically allocate more compute to harder problems, boosting performance without changing models. Open source doesn't offer this yet.

Ecosystem and enterprise support. Polished APIs, safety tuning, compliance certifications, and dedicated support teams still come primarily from proprietary vendors.

For tasks where absolute accuracy justifies a premium -- legal analysis, medical reasoning, high-stakes decision support -- proprietary models remain the right choice. The argument isn't that you should abandon them. It's that you shouldn't use them for everything.

The China Factor

One pattern worth noting: six of the top ten open-source models by capability come from Chinese labs -- Alibaba, Zhipu AI, MiniMax, Xiaomi, Moonshot AI, and DeepSeek. All released under permissive licences (Apache 2.0 or MIT) that allow full commercial use worldwide.

This isn't a geopolitical point. It's a competitive dynamics point. The volume and quality of open-source contributions from Chinese labs has made it structurally impossible for any single proprietary vendor to maintain a durable lead. When half a dozen well-funded labs release frontier-competitive models every quarter, the proprietary advantage erodes by default.

What This Means for Your Business

The data leads to one conclusion: don't tie your agentic workflows to a single provider. Build for flexibility.

Match the model to the task

Not every task needs frontier intelligence. A model performing at 70% of the best handles summarisation, classification, data extraction, and most customer-facing interactions just as well as one at 100% -- at 30-50x lower cost. Smart routing between model tiers can cut your AI spend by 80%+ without any visible drop in quality.

Don't overpay for a brand

The newest frontier model (GPT-5.5 at $17.50/M blended) is now the most expensive non-pro flagship -- and delivers about the same intelligence per dollar as Claude Opus 4.6, which is to say not much. Kimi K2.6 is within 10% of GPT-5.5 in capability at roughly one-sixth the cost ($2.70/M blended), and Gemini 3.1 Pro outperforms Claude Opus at less than half the price. Vendor selection should be driven by data, not habit.

Build for the switch

The best AI model today won't be the best in six months. The cheapest option today will be undercut next quarter. Your infrastructure should let you swap models without rebuilding your applications. The companies that can move fastest will have the biggest advantage.

Take open source seriously

Open source is no longer the budget option you compromise on. Models like Kimi K2.6, Qwen 3.6 Plus, MiniMax-M2.7, and GLM-5.1 are genuine alternatives to GPT-5.4 and Claude 4.6 for the vast majority of use cases. With permissive licensing, you can self-host, fine-tune, and maintain full control over your data -- with no API costs and no vendor dependency.

The Bottom Line

We've seen this pattern before in technology. Proprietary software dominated until Linux proved open-source could match commercial quality. Proprietary databases ruled until PostgreSQL became good enough. Each time, the incumbent maintained a narrowing quality edge while the open alternative closed the gap and won on cost, flexibility, and freedom from lock-in.

AI is following the same trajectory -- compressed into three years instead of fifteen.

Stop choosing one provider. Use them all.

Metir AI is model-agnostic AI infrastructure for agentic workflows. Don't tie your AI to a single provider -- build for flexibility. Try it free.

Data sources: Artificial Analysis Intelligence Index v4.0 (10+ benchmarks), OpenRouter API pricing, official provider pricing pages. Blended prices calculated at 1:1 input-to-output ratio. All data as of April 27, 2026 (post updated to incorporate the GPT-5.5 launch and Moonshot's Kimi K2.6, released April 20, 2026).