If you are running an SMB workload that hits an LLM API directly (a Claude, GPT, or Gemini call from your own code or from n8n / Make.com), the question we keep getting is which provider gives the best price-performance ratio at your volume. The honest answer for sub-1M tokens per month is: at this volume, raw token cost is the wrong tiebreaker. The absolute dollar gap between the cheap tier of each provider is small enough to be invisible on a P&L. The behaviour gap is not.

Below is the actual pricing picture, the reframe of what to compare on, and the recommendation for three specific SMB agent shapes.

What each provider actually costs at the cheap tier in 2026

Pulled from the Anthropic API pricing and Google Gemini API pricing pages on 26 Apr 2026, with OpenRouter’s OpenAI passthrough page as a verifiable proxy for OpenAI’s published rates the same day. All USD per 1M tokens.

Cheap-tier model	Input / 1M tok	Output / 1M tok
Gemini 2.5 Flash-Lite	$0.10	$0.40
GPT-4o-mini	$0.15	$0.60
Gemini 2.5 Flash	$0.30	$2.50
Claude Haiku 4.5	$1.00	$5.00

The number that surprises most operators is the Claude Haiku line. Claude Haiku 4.5 is roughly six to ten times the per-token cost of GPT-4o-mini or Gemini 2.5 Flash-Lite at this tier. That is real and it widens at the output side, where Claude Haiku 4.5’s $5/1M tokens compares to GPT-4o-mini’s $0.60.

Sources: Anthropic API pricing, Google Gemini API pricing, OpenRouter OpenAI page.

What that ratio means at SMB volume

At sub-1M tokens per month, the multiplier is real but the absolute number is small. A typical SMB agentic workload runs in the 200K–800K tokens per month range across all calls. Putting that against the cheap-tier prices:

Workload (input + output tokens / month)	Gemini 2.5 Flash-Lite	GPT-4o-mini	Claude Haiku 4.5
200K total (mixed input/output)	~$0.05	~$0.08	~$0.60
800K total	~$0.20	~$0.30	~$2.40

The Claude Haiku premium at this volume is on the order of $2 per month, not $200. That is below the noise floor of any line item on an SMB P&L, which means price is not the right tiebreaker for the cheap-tier decision at sub-1M tokens.

What is the right tiebreaker

Three things that genuinely differ between the providers at this tier, and that an SMB workload feels:

Tool-use reliability. All three providers ship function-calling at the cheap tier in 2026, but they fail differently. Claude Haiku is the most consistent at returning structured tool calls without hallucinated arguments; this is well-documented in Anthropic’s own tool use docs and reflected in independent tool-calling benchmarks. GPT-4o-mini is close behind and improves faster. Gemini Flash-Lite trails on multi-tool reasoning but holds on single-tool calls.

Instruction-following on long context. Claude Haiku 4.5 and Gemini 2.5 Pro both ship 1M-token context windows; GPT-4o-mini caps at 128K. For SMB workloads that include “summarise this 200-page contract” or “review these 50 customer emails,” only Claude and Gemini operate in-context; GPT-4o-mini requires chunking and re-stitching. The cost of that engineering work is not on the pricing page.

Ecosystem and tooling fit. OpenAI has the broadest tool ecosystem (LangChain, LlamaIndex, every observability vendor). Anthropic has the strongest first-party tooling (Claude Code, Computer Use, Managed Agents) and the deepest MCP integration. Gemini has the cheapest free tier via Google AI Studio for prototyping. The “right” choice tracks which ecosystem your existing stack already lives in.

The 3-shape recommendation

Three SMB agent shapes we see at sub-1M tokens per month, and the cheap-tier model that fits each.

Shape 1: high-frequency classification / triage agent. Inbox triage, support-ticket routing, lead-scoring. High call count, short prompts, structured outputs. Recommend: GPT-4o-mini. Lowest combined cost-plus-engineering at this shape; tool-calling is reliable enough; ecosystem of observability and prompt-management is the deepest.

Shape 2: long-document review agent. Contract review, meeting-transcript summarization, RFP drafting. Low call count, long inputs (50K-500K tokens per call), nuanced outputs. Recommend: Claude Haiku 4.5. The 1M context window and the instruction-following quality on long-document tasks justify the per-token premium at this shape; the absolute dollar cost stays small at sub-1M monthly tokens.

Shape 3: research / synthesis agent feeding a human. Daily news brief, competitor monitoring, market research summaries. Variable inputs, narrative outputs, low call count. Recommend: Gemini 2.5 Flash via the generous free tier or paid Flash at near-zero cost for SMB volume. The cost ratio matters least here because the volume is lowest; Gemini Flash’s research-shape outputs are competitive with the other two; the free tier may carry you indefinitely at this shape.

The two patterns that look reasonable and are not

Pick the cheapest model and route everything to it. The “always Gemini Flash-Lite” approach saves perhaps $1-2 per month at SMB volume and costs you tool-calling reliability on the workloads where Claude or GPT would have been more forgiving. The savings do not justify the support cost when something silently misroutes.

Pick Claude Opus or GPT-5 for safety. The frontier models cost 25-100× more per token than the cheap tier, and at SMB volume their absolute cost is still survivable (~$10-50/month for typical use). The argument against is not cost but fit: most SMB agent workflows do not need frontier reasoning, and using a frontier model on a triage task is paying for capability that does not show up in the output.

What changes this verdict

Cadence on this piece is 30 days because cheap-tier API pricing is the most actively-changing surface in this comparison; both Anthropic and Google have moved cheap-tier prices within 90-day windows in 2025. The three things that would flip the recommendation:

Anthropic ships a sub-$0.50/1M-input cheap-tier model (a true Haiku-class price match against Gemini Flash-Lite and GPT-4o-mini). This would close the only structural cost objection to Claude at SMB volume.
OpenAI extends the 4o-mini context window past 256K. The single biggest gap GPT-4o-mini has at this tier is context length; closing it would expand the workloads where GPT-4o-mini is the right answer.
Google deprecates the generous Gemini API free tier through Google AI Studio. The free-tier latitude is the strongest reason to pick Gemini at this volume; removing it shifts the calculus toward GPT-4o-mini for cost-conscious workloads.

We will re-test against the published pricing pages on or before 26 May 2026. If any of the three preceding conditions has triggered, this claim moves to Partial.